Testing Market Efficiency using Statistical

Arbitrage with Applications to Momentum and

Value Strategies

S. Hogan
a
, R. Jarrow
b
, M. Teo
c*
, M. Warachka
d


a
Credit Suisse First Boston.
b
Johnson Graduate School of Management, Cornell.
c
Singapore Management University and FDO Partners LLC.
d
Singapore Management University.


This Draft: September 2003

Abstract

This paper introduces the concept of statistical arbitrage, a long horizon trading opportunity
that generates a riskless profit and is designed to exploit persistent anomalies. Statistical arbitrage
circumvents the joint hypothesis dilemma of traditional market efficiency tests because its
definition is independent of any equilibrium model and its existence is incompatible with market
efficiency. We provide a methodology to test for statistical arbitrage and then empirically
investigate whether momentum and value trading strategies constitute statistical arbitrage
opportunities. Despite adjusting for transaction costs, the influence of small stocks, margin
requirements, liquidity buffers for the marking-to-market of short-sales, and higher borrowing
rates, we find evidence that these strategies generate statistical arbitrage. Furthermore, their
profitability does not appear to decline over time.


JEL classification: G12; G14
Keywords: Statistical arbitrage; market efficiency; momentum; value



------------------------------------
*
Corresponding author: Address: 168 Mt Auburn St, Cambridge MA 02138. Tel: 617-234-9464.
Fax: 617-234-9478. E-mail address: melvynteo@fdopartners.com

We thank Warren Bailey, Amiyatosh Purnanandam, Bill Schwert, Yiu-Kuen Tse, David
Weinbaum, seminar participants at Cornell and the Singapore Management University, and an
anonymous referee for many helpful comments and suggestions. M. Warachka is grateful for
financial support from the Wharton-SMU Research Center, Singapore Management University.
Credit Suisse First Boston does not necessarily agree with or endorse the content of this article.
1 Introduction
A large number of empirical studies conclude that stock prices appear to contradict the
ecient markets hypothesis. For example, Jegadeesh and Titman (1993) investigate a
trading strategy that buys well-performing stocks and sells poor-performing stocks. They
document average excess returns of 12% per year, where excess returns are dened relative
to a standard capital asset pricing model. Lakonishok, Shleifer, and Vishny (1994) reach
a similar conclusion via buying value and selling glamour stocks identied with variables
such as price earnings ratios, dividends, book to market values, cash ows, and sales growth.
Chan, Jegadeesh, and Lakonishok (1996) conrm the excess returns of portfolios formed on
the basis of past returns and earnings announcements. Many other persistent anomalies
such as those pertaining to dividends, earnings announcements, shares issuances, and share
repurchases have been documented. Shleifer (2000) provides an excellent review of this
literature.
However, all existing studies are compromised by the joint hypothesis problem. Fama
(1998) cautions against rejecting market eciency because previous tests are contingent
upon a specic model for equilibrium returns. Therefore, abnormal returns need not imply
the rejection of market eciency. Instead, they may indicate that the equilibrium model
is misspecied. According to Fama (1998) most long-term anomalies are also sensitive to
the statistical methodology utilized and therefore inferences drawn using long-term returns
(i.e. over 5 years) are circumspect. This paper oers a methodology that addresses Famas
statistical criticisms and circumvents the joint hypothesis problem.
The methodology proposed in this paper is based on the concept of a statistical arbi-
trage. We dene a statistical arbitrage as a long horizon trading opportunity that generates
a riskless prot. As such, statistical arbitrage is a natural extension of the trading strategies
utilized in the existing empirical literature on persistent anomalies. Statistical arbitrage is
the time series analog of the limiting arbitrage opportunity contained in Ross (1976). More
importantly, statistical arbitrage is dened without reference to any equilibrium model, there-
fore, its existence is inconsistent with market equilibrium and, by inference, market eciency
[see Jarrow (1988), chapter 19]. As such, the notion of statistical arbitrage enables the re-
jection of market eciency without invoking the joint hypothesis of an equilibrium model.
1
This paper subsequently provides a statistical test to determine whether a trading strat-
egy constitutes statistical arbitrage based on its incremental trading prots computed over
short time horizons. Our test for statistical arbitrage is intended to replace the standard
intercept test performed in the existing empirical literature on market eciency. Similar
to arbitrage pricing theorys reliance on estimated covariances when determining arbitrage
opportunities, our statistical test utilizes historical data to render a decision on the presence
of statistical arbitrage in the economy. We then apply this new methodology to momen-
tum and value trading strategies. In addition, we also compute their probability of a loss,
providing additional insights into their ability to eventually produce arbitrage prots.
To minimize data mining concerns, the momentum strategies we investigate are mod-
eled after the momentum strategies tested in Jegadeesh and Titman (1993) while the value
strategies are modeled after the contrarian strategies of Lakonishok, Shleifer, and Vishny
(1994). Our sample period is from January 1965 to December 2000 and subsumes that of
both original papers. Our test of statistical arbitrage reveals that of the 16 momentum trad-
ing strategies examined, six produce statistical arbitrage at the 5% level with three more at
the 10% level. The probability of incurring a loss for a momentum strategy that longs the
highest return decile and shorts the lowest return decile based on six months of past returns
and holds that portfolio for twelve months falls below 1% after just 89 months of trading.
Of the 12 value strategies we examine, ve strategies yield statistical arbitrages at the 5%
level. In addition, after just 79 months of trading, the probability of incurring a loss on the
value strategy using the past three years of sales growth with a one year holding period falls
below 1%. In summary, momentum strategies and value strategies provide strong evidence
against the ecient markets hypothesis.
The trading strategies are tested under the assumption that the expected trading prots
are constant over time, which we refer to as the constrained mean (CM) version of statistical
arbitrage. Our rst robustness check evaluates whether the assumption of constant expected
prots is appropriate. We compare the constrained mean (CM) model with an unconstrained
mean (UM) model that incorporates an additional parameter allowing the expected prots
to vary over time. Four independent tests conrm that the CM model is appropriate. First,
the rates of change for expected prots are usually not statistically dierent from zero. Of
the 15 momentum and value strategies which are CM statistical arbitrages at the 10% or 5%
2
level, only one registers a statistically negative rate of decay in its expected prots. Second,
the insample t as measured by the root mean squared error (RMSE) for the CM and UM
models are indistinguishable, indicating there is no advantage to using the more complicated
UM model. Third, the sum of squared normalized residuals are almost identical for both
models. Fourth, likelihood ratio tests cannot reject the null hypothesis that expected prots
are constant over time. Hence we nd that the CM version of statistical arbitrage is the
appropriate test for momentum and value trading strategies. This is because the UM model
spreads the information contained in trading prots over an additional variable without
oering an improved t, thereby weakening the power of the statistical arbitrage test.
To ascertain the robustness of our statistical methodology, we perform a series of sim-
ulations incorporating deviations from our assumed process for incremental trading prots.
Specically, we consider autocorrelation, jumps, and parameter non stationarity. Our sta-
tistical arbitrage test is then applied to these simulated processes. The simulations indicate
that our statistical test is biased towards accepting the null hypothesis of no statistical ar-
bitrage, and thus the acceptance of market eciency. Consequently, instances of statistical
arbitrage found herein are hard to reconcile with market eciency.
A possible concern is that market ineciencies may disappear after including transactions
costs. To investigate this issue, we compute the turnover of each portfolio and combine these
results with the estimated round trip transactions costs of Chan and Lakonishok (1997). We
then adjust the incremental prots downward by these transactions costs. Besides incorpo-
rating transaction costs, we also consider four additional market frictions as in Alexander
(2000). We study the impact of requiring margin for both long and short positions as well
as reduced proceeds from the interest earned on the margin account. We also jointly assume
an additional liquidity buer for the short position and borrowing rates that are higher than
lending rates. In this realistic trading environment, our conclusions regarding the presence
of statistical arbitrage remain unchanged as 9 of the 11 statistical arbitrages at the 5% level
remain statistical arbitrages at the 5% level. This suggests that the statistical arbitrages
represent feasible and attractive trading opportunities for market participants.
A common feature of stock market anomaly studies is that small stocks are often less e-
cient than large stocks (Hong, Lim, and Stein, 2000; Mitchell and Staord, 2000). Moreover,
since small stocks may require greater transactions costs and are less liquid, we investigate
3
whether our results survive their exclusion. Even after removing stocks with market equity
below the 50% NYSE percentile from our sample, all but one of the momentum portfolios
and 7 out of 12 value portfolios still test positively for statistical arbitrage at the 5% level.
Thus, our previous results do not appear to be driven by small stocks.
Lastly, we investigate the dependence of our results on the sequence of observed returns.
Using a bootstrap procedure, we generate multiple trading prot sequences from the observed
time series and apply our test for statistical arbitrage. This robustness test re-arms the
presence of statistical arbitrage in momentum and value strategies.
As an interesting application of the statistical arbitrage methodology, we investigate the
size eect. We test the Fama and French (1993) SMB factor for statistical arbitrage. The t-
statistic for the SMB factor over our sample period suggests that one will be rewarded with
positive expected prots when investing in a strategy that longs small stocks and shorts
large stocks. However we nd that this SMB based trading strategy does not constitute a
statistical arbitrage, conrming the popular belief that the size eect disappeared after the
early 1980s.
The remainder of this paper is organized as follows. Section 2 introduces the concept of
statistical arbitrage and its relation to market eciency. Section 3 derives statistical tests
for the existence of statistical arbitrage. The data used in our empirical study is described in
Section 4 while Section 5 presents our empirical results. Section 6 describes various tests to
gauge robustness, while Section 7 applies the statistical arbitrage methodology to the testing
of the size eect. Section 8 concludes and oers avenues for further research.
2 Statistical arbitrage and market eciency
This section introduces the notion of a statistical arbitrage and its relation to market e-
ciency. Traded in the economy are a stock (or portfolio) S
t
and money market account B
t
initialized at a dollar (B
0
= 1).
Trading strategies are central to the notion of statistical arbitrage. As in the standard
option pricing literature, let the stochastic processes (x(t), y(t) : t 0) represent a zero
initial cost, self-nancing trading strategy involving x(t) units of a stock and y(t) units
of a money market account at time t. This trading strategy is formulated using only
4
available information such as past returns, rm sizes, earnings announcements, market versus
book values, sales growth rates, or macroeconomic conditions. These trading strategies, by
denition, must have zero initial cost, x(0)S
0
+ y(0) = 0. For emphasis, the stock could
represent a (zero cost) self-nancing portfolio consisting of long and short positions in various
risky assets.
Although models of market equilibrium may assist in identifying possible trading strate-
gies, a model of market equilibrium is not required. Let the process V (t) denote the cumu-
lative trading prots at time t that are generated by such a trading strategy (x(t) : t 0).
1
In the following analysis, it will be necessary to work with the discounted value of the cu-
mulative trading prots dened as v(t) = V (t)/B
t
. To illustrate these concepts and various
characteristics of typical trading strategies, we consider a few examples.
Example 1 Consider the standard Black-Scholes economy with a non-dividend paying stock
price S
t
that evolves according to
S
t
= S
0
e
t
2
t/2+Wt
and a money market account B
t
= e
rt
where ,
2
, r are non-negative constants, > r, and
W
t
is a standard Brownian motion. Consider the self-nancing trading strategy that consists
of buying and holding 1 unit of the stock, nanced with the money market account. The
value of this portfolio at time t is
V (t) = 1 S
t
S
0
e
rt
= S
0

e
[
2
/2]t+Wt
e
rt

representing the cumulative trading prots with an initial value V

0
= 0. The expectation and
variance
2
of the cumulative discounted trading prots both approach innity as t ,
E
P
[v(t)] = S
0

e
(r)t
1

and V ar
P
[v(t)] = S
2
0
e
2(r)t

2
t
1

.
In addition, the time-averaged variance
V ar
P
[v(t)]
t
also converges to as t , the signi-
cance of which will become apparent when statistical arbitrage is dened.
1
The process V (t) is dollar denominated and it is neither a cumulative excess return nor a cumulative
residual with respect to an equilibrium model.
2
See Casella and Berger (1990) page 628 for the variance of a lognormal random variable.
5
As a visual illustration, the discounted cumulative payos after one, two, three, and ve
years are simulated for the above trading strategy which nances the purchase of stock by
borrowing from the money market account. A total of 10,000 simulations are conducted
with histograms of the payos provided in Figure 1. Observe that the distribution is right-
skewed, consistent with its lognormal distribution, and appears to shift to the right over time.
Although the mean increases, both the variance and the time-averaged variance increase over
time, with the second property
3
the focus of two subsequent examples. Hence Example 1 does
not constitute a statistical arbitrage. Note also that a trading strategy that only buys stock,
without borrowing, would neither be self-nancing nor would have a declining time-averaged
variance.
Many instances of premiums or excess returns in the anomalies literature are generated
by portfolios that are long and short portfolios of stocks formed on the basis of certain
cross sectional characteristics. The next example involves arithmetic rather than geometric
Brownian motion, a reasonable process for a portfolio of long and short positions that is not
a limited liability asset.
Example 2 Consider the discounted cumulative value of a trading strategy v(t) that evolves
according to
v(t) = t +W
t
with v(0) = 0. The expectation and variance of the cumulative discounted trading prots are
both unbounded as t
E
P
[v(t)] = t and V ar
P
[v(t)] =
2
t
while the time-averaged variance is nite and equals
2
.
The previous examples illustrate trading strategies that are not long horizon excess
prot opportunities. This is because as the trading strategies expected prots increase
3
Incidentally, the probability of a loss is easily computed by evaluating the terminal value of the dis-
counted cumulative trading prots for each time horizon after every simulated path, to complement the
visual presentation given in the histograms. This property is discussed later in more detail as the examples
are intended to emphasize the strategys time-averaged variance.
6
across time, there is a corresponding increase in the trading prots variance. In particular,
the time-averaged variance does not disappear. In contrast, the following example captures
the intuitive notion of a statistical arbitrage.
Example 3 Let us revisit the buy and hold trading strategy underlying Example 2. Suppose
that the discounted trading prots over an arbitrary intermediate trading interval [t
k1
, t
k
]
can be represented as
v(t
k
) v(t
k1
) = +z
k
where , > 0 and z
k
are i.i.d. random variables with zero mean and variance 1/k. This
trading strategy has positive expected discounted prots over every interval (), but with
random noise (z
k
) appended. The variance of the noise is decreasing over time. Once
again, v(t
0
) = 0 and the cumulative discounted trading prots at time t
n
equals
v(t
n
) =
n

k=1
[v(t
k
) v(t
k1
)] = n +
n

k=1
z
k
.
Observe that E
P
[v(t
n
)] = n and V ar
P
[v(t
n
)] =
2

n
k=1
1
k
both converge to innity.
However,
V ar
P
[v(tn)]
n
=
2
!
n
k=1
1
k
n
0 as n .
Over time, analogous to cross-sectional diversication in Ross APT, the random noise
in the trading strategy of Example 3 is diversiable. Indeed, this trading strategy oers
a positive discounted expected prot with a time-averaged variance approaching zero and
captures the intuition behind statistical arbitrage. Given these insights, we now formalize
the notion of a statistical arbitrage.
Denition 4 A statistical arbitrage is a zero initial cost, self-nancing trading strategy
(x(t) : t 0) with cumulative discounted value v(t) such that:
1. v(0) = 0
2. lim
t
E
P
[v(t)] > 0
3. lim
t
P(v(t) < 0) = 0, and
4. lim
t
V ar
P
[v(t)]
t
= 0 if P(v(t) < 0) > 0 t < .
7
By denition, a statistical arbitrage satises four conditions (i) it is a zero initial cost
(v(0) = 0) self-nancing trading strategy, that in the limit has (ii) positive expected dis-
counted prots, (iii) a probability of a loss converging to zero, and (iv) a time-averaged
variance converging to zero if the probability of a loss does not become zero in nite time.
That is, the fourth condition only applies when there always exists a positive probability of
losing money. Otherwise, if P(v(t) < 0) = 0 for all t T for some T < , then a standard
arbitrage opportunity is available. In economic terms, the fourth condition implies that
a statistical arbitrage opportunity eventually produces riskless incremental prot, with an
associated Sharpe ratio increasing monotonically through time. This becomes apparent
in the next section when the convergence rate to arbitrage is studied. This is consistent
with the variance of the trading strategy increasing towards innity with time, but with
a growth rate less than linear. The fourth property of statistical arbitrage essentially
mitigates the concerns of Shleifer and Vishny (1997) with respect to risky arbitrage. Ad-
ditional justication for the fourth condition is given in Appendix A.1. Overall, the ability
of a trading strategy to reduce its time-average variance over time
4
is essential to generate
statistical arbitrage. Returning to Example 1, we observe that the buy and hold strategy
involving the purchase of a stock nanced with the money market account is not capable of
generating statistical arbitrage. The same conclusion applies to Example 2. Only Example
3 is consistent with statistical arbitrage.
A standard arbitrage opportunity is a special case of statistical arbitrage. Indeed, a
standard arbitrage has V (0) = 0 where there exists a nite time T such that P(V (t) > 0) > 0
and P(V (t) 0) = 1 for t T. To transform the standard arbitrage opportunity into an
innite horizon self-nancing trading strategy, we just invest the proceeds at time T into the
money market account, i.e. V (s) = V (T)
Bs
B
T
for s T. Note that v(s) = V (T)
Bs
B
T
1
Bs
= v(T).
Then, lim
s
E
P
[v(s)] = E
P
[v(T)] > 0 which satises condition 2 and lim
s
P(v(s) < 0) =
P(v(T) < 0) = 0 which satises condition 3 and implies that condition 4 is not applicable.
However, this papers intent is to construct a statistical test for evaluating market eciency
with regards to persistent anomalies, trading strategies whose probability of losing money is
always positive at a given nite point in time.
4
The self-nancing condition alters the composition of the portfolio over time, creating an interaction
between the self-nancing and the decreasing time-averaged variance properties.
8
This denition is similar to the limiting arbitrage opportunity used to construct Ross
APT (1976). The dierence between the two concepts is that a statistical arbitrage is a
limiting condition across time, while Ross APT is a cross-sectional limit at a point in time.
This dierence necessitates discounting by the money market account and the normalization
by time in condition 4. Therefore, just as Ross APT is appropriate in an economy with a
large number of assets, our methodology is appropriate for long time horizons. Investors
with long but nite time horizons view statistical arbitrage opportunities as too good to
pass up as they oer positive expected discounted prot, variance (per unit time) that
becomes arbitrarily small, and decreasing risk of a loss. Indeed, although statistical arbitrage
is dened over an innite time horizon, there exists a nite timepoint t

such that the

probability of a loss is arbitrarily small, P(v(t

) < 0) = . Thus, comparing an arbitrage

opportunity at time t

having the property P(v(t

) < 0) = 0 with a statistical arbitrage

opportunity reveals they are only separated by an probability of a loss.
It is well known that arbitrage opportunities are incompatible with an ecient market, see
Jensen (1978) for a general discussion
5
. Tests of market eciency in the options market based
on the absence of arbitrage are numerous. For example, Kamara and Miller (1995) study
violations of put-call parity, while Bodurtha and Courtadon (1986) investigate the eciency
of foreign currency options. Tests for statistical arbitrage rejecting market eciency can
be viewed as an extension of this literature. Indeed, statistical arbitrage rejects the market
as being in any economic equilibrium, an important prerequisite for an ecient market (see
Jarrow (1988), chapter 19). A statistical arbitrage (analogous to an arbitrage opportunity in
Ross APT) will induce investors to engage in trading, and consequently prevent the market
from reaching an equilibrium. Even a single trader operating under a nite time horizon and
willing to pursue a statistical arbitrage is sucient to conclude these trading opportunities
reject market eciency.
Our idea of extending standard arbitrage to its innite horizon counterpart is related
to a literature extending simple arbitrage in alternative ways. In the context of incomplete
markets, Cochrane and Sa a-Requejo (2000) invoke Sharpe ratios to nd asset prices that
5
Jensen (1978) states that the existence of economic trading prots (dened therein on page 96) reject an
ecient market. An arbitrage opportunity (and a statistical arbitrage) satises his denition of economic
trading prots.
9
are good deals to investors while Bernardo and Ledoit (2000) exclude investments whose
maximum gain-loss ratios are too attractive. Both of these approaches also investigate trading
opportunities that generalize the denition of arbitrage without specifying a particular model
of market equilibrium. Carr, Geman, and Madan (2001) introduce the concept of acceptable
opportunities which would be executed by any reasonable investor, implying their existence
is incompatible with an ecient market. This paper introduces statistical arbitrage as a
similar generalization of arbitrage.
This notion of statistical arbitrage can also be extended to study the traditional weak- and
strong-form tests of market eciency. Indeed, in constructing the trading strategies, weak-
form statistical arbitrage would utilize only publicly available information, while strong-form
statistical arbitrage would utilize all available information.
Furthermore, statistical arbitrage requires no information beyond the traditional inter-
cept test. Indeed, portfolio returns from long and short positions are easily converted into
dollar denominated trading prots with gains and losses accruing through time according to
the riskfree rate, as shown in our later empirical analysis.
An important issue regarding the implementation of our trading strategies is the exclusion
of doubling strategies (see Due (2001)). In our later empirical work, $1 is kept constant in
the risky asset (long minus short portfolio position) over time. Investing the trading prots,
which have positive expectation, in the riskless money market account serves to reduce the
trading strategys time-averaged variance. Appendix A.2 proves that doubling strategies are
incompatible with the second and fourth properties of statistical arbitrage in Denition 4.
Moreover, our empirical study does not terminate trading at the rst instance of a positive
cumulative trading prot. Dybvig and Huang (1988) prove that a non-negative lower bound
on wealth is sucient to preclude arbitrage opportunities of the doubling persuasion. To
appeal to this economically sensible restriction, we record the maximum loss for trading
strategies that generate statistical arbitrage in our empirical applications.
There are several dierences between the statistical arbitrage test and the traditional
tests of market eciency using a risk-adjusted alpha. First and most important, the risk-
adjustment process for returns requires an underlying model of market equilibrium. This
makes alpha, as a test of market eciency, subject to the joint hypothesis critique. In
contrast, the test for statistical arbitrage test does not require a model of market equilibrium
10
to generate risk-adjusted or excess returns. Instead, dollar denominated incremental prots
from a self-nancing trading strategy are analyzed when testing for statistical arbitrage.
Second, a linear factor model is typically assumed in the risk-adjustment process (e.g. Fama-
French, 1993; Carhart, 1997). The statistical arbitrage test can be applied to any asset,
including those that are not priced using linear factor models such as derivatives. Third,
the alpha test on the mean is unable to detect the presence of statistical arbitrage. As
demonstrated in the next section, the statistical arbitrage test does not reduce to a t-ratio
test. Such a reduction would require the rate of change in the trading prots volatility to
be zero and violate the crucial fourth condition of statistical arbitrage. Fourth, a statistical
arbitrage requires the time-averaged variance of a strategy to decline while the alpha test
has no such condition.
3 Tests for statistical arbitrage
In this section, we derive statistical tests for determining whether a trading strategy yields
statistical arbitrage. Two classes of statistical tests are considered: constrained mean (CM)
tests and unconstrained mean (UM) tests. This distinction is important when we investigate
momentum and value strategies in the later sections.
Observe that our proposed framework is an empirical methodology that oers an alterna-
tive to the standard risk-adjusted intercept test by circumventing the required specication
of an equilibrium model for returns. However, as an empirical methodology, the statistical
test is conned to testing whether statistical arbitrage opportunities existed ex-post in the
sample period rather than ex-ante.
3.1 The Statistical Arbitrage Testing Methodology
To test for statistical arbitrage, a time series of dollar denominated discounted cumulative
trading prots v(t
1
), v(t
2
), . . . , v(t
n
) generated by a trading strategy are analyzed. For a given
trading strategy, let v
i
= v(t
i
) v(t
i1
) denote increments
6
of the discounted cumulative
6
Trading prots derived from portfolios are well represented by a normal distribution since the impact of
idiosyncratic jumps are mitigated.
11
trading prot measured at equidistant time points t
i
t
i1
= with t
i
= i .
Our test for statistical arbitrage requires an assumed process for the evolution of incre-
mental trading prots. At rst, for simplicity, we impose the assumption that the trading
prots have independent increments. This assumption is later relaxed but empirical results
conrm that it does not alter our conclusions regarding market eciency. For greater gen-
erality in presenting our results, we also allow the expected prots of the trading strategy to
vary over time. Thus, we begin our analysis with the unconstrained mean (UM) model. The
quantity denoting the time between equidistant increments may be set to one without any
loss of generality with t
i
understood as being i.
Assumption 1 Let the discounted incremental trading prots
7
satisfy
v
i
= i

+i

z
i
(1)
for i = 1, 2, . . . , n where z
i
are i.i.d N(0, 1) random variables with z
0
= 0. Note that both
v(t
0
) and v
0
are zero.
A Taylor series expansion reveals that our formulation encompasses linear, and potentially
quadratic specications, depending on the magnitude of and . More importantly, these
specications account for the changing composition of a statistical arbitrage portfolio as
trading prots are invested in the riskfree account, causing proportionally less to be invested
in the risky position over time. In particular, consider an expansion of i

with respect to i
which yields (i +1)

= i

1 +

i
+
(1)
2i
2

plus higher order terms. For < 0, an important

condition of statistical arbitrage shown later in this section, the marginal decrease in the
portfolios volatility itself declines over time. This property is consistent with the declining
marginal impact of the riskfree asset. In our empirical implementation, a protable trading
strategy generates trading prots V (t) (undiscounted) that are invested in the riskfree asset
while a constant amount of $1 is invested in both the long and short portfolios over time.
For statistical arbitrage opportunities, the fraction of the portfolio invested in the risky
asset (long - short portfolio) equals 1 divided by 1 plus the trading prots,
1
1+V (t)
, a convex
7
Note that Assumption 1 implies that condition 4 in the denition of statistical arbitrage is relevant.
It excludes standard arbitrage opportunities, whose existence can be tested via alternative and well-known
procedures.
12
decreasing function when trading prots increase over time. This feature parallels the i

(and the associated specication for the mean) component of the volatility function.
Moreover, need not be a constant. For example, volatility (t) could evolve as a
GARCH process. This enhancement is not considered in our later empirical work although
the main theorem of this section remains valid with this generalization.
The expectation and variance of the discounted incremental trading prots in equation
(1) are E[v
i
] = i

and V ar[v
i
] =
2
i
2
. Observe that for < 0, the variance of v
i
decreases over time, which ultimately satises the fourth condition of statistical arbitrage
as seen later. Also note that the dynamics of the discounted incremental trading prot
in equation (1) with < 0 and > 0 does not advocate a trading strategy based on
waiting for the volatility to decline before investing. Instead, it is optimal for investors to
immediately begin trading and earn a positive expected prot while enjoying the benet of
decreasing time-averaged variance. A bootstrap procedure in Section 6 conrms that the
observed sequence of portfolio returns is immaterial to whether the trading strategies under
consideration constitute statistical arbitrage opportunities.
The discounted cumulative trading prots generated by the trading strategy are
v(t
n
) =
n

i=1
v
i
d
N

i=1
i

,
2
n

i=1
i
2

(2)
while the log likelihood function for the increments in equation (1) equals
log L(,
2
, , |v) =
1
2
n

i=1
log

2
i
2

1
2
2
n

i=1
1
i
2

v
i
i

2
(3)
allowing maximum likelihood estimation to generate the four required parameters. The score
equations required to solve for the four parameters are found
8
in the next lemma.
Lemma 5 Parameter estimates are obtained by solving the following four equations for the
8
The property

x
t
x
= log(t)t
x
is used repeatedly when solving for

and

.
13
four unknown parameters.
log L(,
2
, , |v)

: =

n
i=1
v
i
i

n
i=1
i
2(

)
(4)
log L(,
2
, , |v)

2
:
2
=
1
n
n

i=1
1
i
2

v
i
i

2
(5)
log L(,
2
, , |v)

:
n

i=1
v
i
log(i)i

=
n

i=1
log(i)i
2(

)
(6)
log L(,
2
, , |v)

:
2
n

i=1
log(i) =
n

i=1
log(i)
i
2

v
i
i

2
(7)
The consistency and eciency of these Maximum Likelihood Estimates (MLE) are both
well-established, see Chapter 7 of Casella and Berger (1990). As expected, for = 0 and
= 0, equations (4) and (5) in the previous lemma reduce to the standard normal MLE
estimators
=
1
n
n

i=1
v
i
and
2
=
1
n
n

i=1
(v
i
)
2
.
The functional form of the coecients in equation (1) allows for a large
9
variety of potential
changes to the mean and variance. Moreover, as demonstrated later in this paper, misspeci-
fying the incremental trading prot process only serves to increase the likelihood of accepting
the null hypothesis of no statistical arbitrage. Therefore, the existence of statistical arbi-
trage, the subject of the following theorem, is robust to criticisms regarding the specied
stochastic process.
Theorem 6 A trading strategy generates a statistical arbitrage with 1 percent condence
9
The specication of the coecients implicitly accounts for the usual rescaling of large positive numbers,
such as trade volume, using a log transformation. A Taylor series expansion of i

with respect to evaluated

at = 0 equals i

= 1 + ln(i) +
1
2
(ln(i))
2

2
plus higher order terms. Thus, our formulation conforms with
the standard empirical practice of taking logs for large positive variables as the index i reaches values over
400 (number of months) in our later empirical implementation.
14
if the following conditions are satised:
H1: > 0 ,
H2:

< 0 ,
H3:

> max

1
2
, 1

,
with the sum of the individual p-values forming an upper bound for the tests Type I error.
Therefore, the sum of the p-values associated with the individual hypotheses must be below
to conclude that a trading strategy generates statistical arbitrage.
Proof. In order to satisfy the fourth condition of statistical arbitrage, equation (2)
implies the parameter must be negative to ensure

2
!
n
i=1
i
2
n
0 while the second condition
requires the parameter to be positive. Any value of ensures

n
i=1
i

> 0 provided > 0.

However, the convergence of P(v(t) < 0) to zero requires

n
i=1
i

n
i=1
i
2

with details in Appendix A.3. The leading order term of the sum in the numerator (using
left endpoints)
n

i=1
i

n
1
s

ds =
n
1+
1 +

1
1+
1 +
divided by the square root of the sum in the denominator (using right endpoints)
n

i=1
i
2

n
1
(s + 1)
2
ds =
(n + 1)
2+1
2 + 1

2
2+1
2 + 1
equals
n
+1
(n+1)
+
1
2
. Therefore, the issue of the tail probability converging to zero is whether
n
+
1
2
converges to innity which is the case when >
1
2
. The leading order approx-
imation is only valid
10
in the numerator if > 1. Therefore, with a negligible loss of the
endpoint 1, the third condition of statistical arbitrage requires > max{
1
2
, 1}.
10
For = 1,

n
i=1
i
1
is a harmonic series which diverges to innity with integral representation

n
1
s
1
ds = ln(n). Therefore, = 1 is the critical value which ensures the numerator approaches in-
nity, a condition required to ensure the probability of a loss approaches zero. Values of smaller than -1
result in the numerator converging to a nite number.
15
The three parameters are tested individually with the Bonferroni
11
inequality, which stip-
ulates that the sum of the p-values for the individual tests becomes the upper bound for the
Type I error of the joint hypothesis test.
12
Standard errors for the hypothesis tests in Theo-
rem 6 are extracted from the Hessian matrix to produce t-statistics and their corresponding
p-values.
As demonstrated above in Theorem 6, based on the distribution of discounted cumulative
trading prots found in equation (2), the second and third conditions of statistical arbitrage
require > 0 and >
1
2
, while the fourth condition requires < 0. The condition
>
1
2
ensures the probability of a loss converges to zero. The parameters may be
negative, indicating that expected trading prots are decreasing over time, provided they
are not negative enough to prevent convergence to arbitrage. Thus, the third hypothesis is
equivalent to a test of long run market eciency. Observe that the existence of statistical
arbitrage is not rejected if equals zero, in contrast to the situation when either or
are zero. This leads to a constrained mean test for statistical arbitrage given in the next
subsection.
Furthermore, it is imperative to recognize that applying a single t-statistic to trading
prots is not a valid test for statistical arbitrage. Assume and are both zero, implying the
cumulative trading prot v(t) in equation (2) is distributed N(n,
2
n). Given an estimate

2
=
1
n

n
i=1
(v
i
)
2
, Theorem 6 reduces to a t-test, with n 1 degrees of freedom, of
=
1
n

n
i=1
v
i
being statistically positive
> t
n1,

2
n
.
However, this single t-test is unable to test for statistical arbitrage since the fourth property
11
P (
n
i=1
A
i
)

n
i=1
P(A
i
), Casella and Berger (1990) page 11. The Bonferroni inequality sums the
probabilities of the events in the union without subtracting the probability of their intersection to obtain an
upper bound.
12
There is no uniformly most powerful test (Denition 8.3.5 Casella and Berger (1990) page 365) as the
conditions for applying the Neyman-Pearson Lemma (Corollary 8.3.2. Casella and Berger (1990) page 368)
on composite hypotheses are not satised by the non monotonic likelihood function in equation (3). A
likelihood ratio test (Denition 8.2.1 Casella and Berger (1990) page 347) would be biased towards over-
rejecting the null hypothesis of no statistical arbitrage since all three conditions in Theorem 6 have to be
satised in order to reject the null.
16
of statistical arbitrage cannot be veried. In summary, a single t-test is limited to only testing
the hypothesis of positive expected trading prots and cannot detect statistical arbitrage.
Besides the existence of statistical arbitrage, the rate of convergence to arbitrage is also
interesting and the subject of the next theorem which states that convergence of the tail
probability P (v(t
n
) < 0) is faster than exponential. Thus, our introduction of an innite time
horizon is actually less onerous than it appears, as investors quickly realize an investment
opportunity that is close to arbitrage.
Theorem 7 The rate of convergence from statistical arbitrage to arbitrage is faster than
exponential. Specically, the convergence rate is faster than
O

exp

n
i=1
i

2
2
2

n
i=1
i
2

.
The proof is contained in Appendix A.3. Observe that the rate of convergence depends
on the squared Sharpe ratio

2

2
as well as the rates of change and . Trading strategies
may be evaluated by the convergence rate of P (v(t
n
) < 0) to zero and the desirability of
trading strategies ranked according to this function.
Besides implying the Sharpe ratio increases towards innity, we also consider an investor
with log utility presented with a statistical arbitrage opportunity. As proved below, such
an investor would nd it desirable to pursue a statistical arbitrage opportunity. Further-
more, their demand for the statistical arbitrage opportunity is unbounded, as is the case for
standard arbitrage opportunities.
For any level of exogenous wealth
13
W > 0 in the future, there exists a time horizon t

such that for all t

n
t

,
E [U(W +v(t
n
))] > E [U(W)] > 0 .
Intuitively, as with a standard arbitrage opportunity, the investor increases their expected
utility by pursuing a statistical arbitrage opportunity. Indeed, taking an innitely large
position in the zero cost self-nancing asset is possible. Hence, the economy cannot be in
13
We suppress the time dependence of W for simplicity as it is not the focus of our present analysis.
This level of wealth may be random, in which case our proof would rst condition on this amount before
continuing in its present form.
17
equilibrium if such trading strategies exist. We consider the case of log utility and prove
that
E[U(W +v(t
n
)) U(W)] = E

ln

1 +

W
v(t
n
)

> 0 > 0 (8)

where the equality stems from the identity ln(x+y)ln(x) = ln

x+y
x

= ln

1 +
y
x

. Equation
(8) may be expressed in terms of v(t
n
)s normal distribution v(t
n
)
d
N

n
i=1
i

,
2

n
i=1
i
2

.
Dene M
tn
as 1+

W
v(t
n
) to express the inequality in equation (8) as E[ln(M
tn
)] > 0. Under
the conditions of statistical arbitrage in Theorem 6, since

W
is positive, Appendix A.3 proves
that P(M
tn
1 < 0) = P

W
v(t
n
) < 0

= P(v(t
n
) < 0) 0. Therefore, equation (8) is
satised for all as there exists a t

such that for all t

n
t

, P(M
tn
> 1) = 1 for > 0
arbitrarily small, implying E[ln(M
tn
)] > 0 for the same set of t
n
.
3.2 Unconstrained versus constrained mean statistical arbitrage
The results of this section remain valid when the parameter is set to zero, implying the
trading strategys expected prots are constant over time. In this constrained mean model,
special cases of Lemma 5 as well as Theorems 6 and 7 immediately result. We refer to this
second test as the constrained mean (CM) test for statistical arbitrage to dierentiate it
from the unconstrained mean (UM) test in Theorem 6.
We begin our empirical analysis with the CM test and later verify that this formulation
is appropriate. A variety of robustness checks in Section 6 support the usage of the CM
model versus the enhanced UM specication in the context of momentum and value trading
strategies. With the parameter = 0, the process is equation (1) reduces to
v
i
= +i

z
i
.
The corresponding joint hypothesis test for statistical arbitrage is:
H1: > 0 ,
H2: < 0 ,
while the probability of loss after n periods (see Appendix A.3) is equal to
Probability of a Loss (after n periods) = N

n
i=1
i
2

(9)
where N() denotes the cumulative standard normal distribution.
18
4 Data
Our sample period starts on January 1965 and ends in December 2000. Monthly equity
returns data are derived from the Center for Research in Security Prices at the University of
Chicago (CRSP). Our analysis covers all stocks traded on the NYSE, AMEX, and NASDAQ
that are ordinary common shares (CRSP sharecodes 10 and 11), excluding ADRs, SBIs,
certicates, units, REITs, closed-end funds, companies incorporated outside the U.S., and
Americus Trust Components.
The stock characteristics data are derived from the CRSP-COMPUSTAT intersection.
The stock characteristics we focus on include book-to-market equity, cash ow-to-price ratio,
earnings-to-price ratio and annual sales growth. To calculate book-to-market equity, book
value per share is taken from the CRSP / COMPUSTAT price, dividends, and earnings
database. We treat all negative book values as missing. We take the sum of COMPUSTAT
data item 123 (Income before extraordinary items (SCF)) and data item 125 (Depreciation
and amortization (SCF)) as cash ow. Only data 123 item is used to calculate cash ow if
data 125 item is missing. To compute earnings, we draw on COMPUSTAT data item 58
[Earnings per share (Basic) excluding extraordinary items] and to compute sales we utilize
COMPUSTAT data item 12 [sales (net)]. Also, all prices and common shares outstanding
numbers employed in the calculation of the ratios are computed at the end of the year.
To ensure that accounting variables are known before hand and to accommodate the
variation in scal year ends among rms, sorting on stock characteristics is performed in
July of year t based on accounting information from year t-1. Hence, following Fama and
French (1993), to construct the book-to-market deciles from July 1st of year t to June 30th
of year t+1, stocks are sorted into deciles based on their book-to-market equity (BE/ME),
where book equity is book equity in the scal year ending in year t-1 and market equity
is calculated in December of year t-1. Similarly, to construct the cash ow-to-price deciles
from July 1st of year t to June 30th of year t+1, stocks are sorted into deciles based on
their cash ow-to-price, where cash ow is cash ow in the scal year ending in year t-1 and
price is the closing price in December of year t-1. Earnings-to-price is calculated in a similar
fashion. All portfolios are rebalanced every month as some rms disappear from the sample
over the 12 month period.
19
An investment of $1 is maintained in the portfolios at all times. The self-nancing con-
dition is enforced by investing (borrowing) trading prots (losses) generated by the various
trading strategies in the riskfree asset. Riskfree rate data are obtained from Kenneth Frenchs
website.
5 An application to momentum and value strategies
This section presents the results of the tests for statistical arbitrage performed on momentum
and value strategies under the assumption that expected incremental trading prots are
constant over time. Two hypotheses are jointly tested. First, the incremental prots from
the strategy must be statistically greater than zero and second, the time-averaged variance
of the strategy must decline to zero as time approaches innity. Table 1 contains summary
statistics for the incremental trading prots of the momentum and value trading strategies
under investigation.
In terms of our previous notation in Section 2, the cumulative trading prot V (t
i
) is a
function of its past value V (t
i1
) and the current prot earned on the $1 invested in the
long and short positions. Specically, the trading strategy x(t) is a 2 by 1 vector, [1, 1]
T
across time while the amount invested in the money market, y(t), simply equals the previous
cumulative trading prot, V (t
i1
). Let L(t
i
) and S(t
i
) represent the return of the long and
short portfolio respectively. With $1 in the these portfolios, the current trading prot at
period t
i
equals 1 + L(t
i
) (1 +S(t
i
)) = L(t
i
) S(t
i
). The previous trading prot V (t
i1
)
accumulates at the riskfree rate r(t
i
) to yield V (t
i
) = L(t
i
)S(t
i
)+V (t
i1
)[1+r(t
i1
)]. This
cumulative trading prot V (t
i
) is then discounted by B(t
i
) = exp

i
j=1
r(t
j
)

to produce
v(t
i
), whose rst dierence v(t
i
) is analyzed for statistical arbitrage.
5.1 Momentum strategies
The momentum strategies we focus on are modeled directly after the strategies tested in
Jegadeesh and Titman (1993). Specically we examine strategies that long stocks in the top
return decile and short stocks in the bottom return decile. These strategies are based on
formation periods of 3, 6, 9, and 12 months and are held for 3, 6, 9, or 12 months.
20
For example, consider the construction of a momentum strategy with a formation period
of three months, and a holding period of six months (MOM 3/6). Each month, we look
back three months to nd the top return decile stocks and the bottom return decile stocks.
Next, we long the top return decile, short the bottom return decile, and hold this portfolio
for six months. The portfolio is rebalanced monthly to account for stocks that drop out of
the database.
Given the possible permutations of formation and holding periods, we examine 16 momen-
tum strategies in total. While there are certainly momentum strategies with other formation
and holding period combinations that one can examine, doing so might expose our results to
a data snooping criticism. Moreover, all trading strategies originate from the same dataset,
as described in Section 4. Thus, statistical tests applied to dierent trading strategies are
not independent, an issue that permeates existing empirical studies. Even if one ignores this
dependence, out of the 28 trading strategies we test, 2 trading strategies would incorrectly
test positive for statistical arbitrage at the 7% level (2/28), slightly above the 5% level we
consider sucient to reject market eciency. Consequently, Type I error in conjunction with
the possibility of dependence amongst the trading strategies may explain a portion of our
later empirical ndings.
The results of our tests for statistical arbitrage are shown in Table 2. We also present t-
ratio tests on the expected prots of the portfolios for comparison. Consistent with Jegadeesh
and Titman (1993), the portfolios expected prots are almost always statistically greater
than zero.
14
Figure 2 plots and contrasts the cumulative trading prots of two momentum
strategies: MOM 6/9 and MOM 12/12. Clearly, MOM 6/9 is preferable to MOM 12/12 even
though they both produce positive expected prots.
The statistical arbitrage results from Table 2 conrms the presence of market ineciency
based on several momentum strategies. For 14 of the 16 portfolios, the point estimate for
the mean () is greater than zero, and the point estimate for the growth rate of the variance
() is less than zero, consistent with statistical arbitrage. Furthermore, at the 5% level, 13
out of 16 momentum trading strategies have statistically positive estimates for . However,
14
It is important to note that the mean incremental prot, , is related but not identical to the usual
mean returns from the trading strategy since is a dollar denominated quantity derived from a self-nancing
trading strategy.
21
the requirement that the time-averaged variance of the incremental prots decline over time
appears harder to satisfy. Only 6 of the 16 trading strategies have signicantly negative
estimates for at the 5% level. Nonetheless, the same 6 trading strategies are able to
generate statistical arbitrage at the 5% level. Another 3 trading strategies yield statistical
arbitrage at the 10% level. In short, roughly half of the momentum strategies examined
converge to riskless arbitrages with decreasing time-averaged variances. This nding is hard
to reconcile with the notion of market eciency.
The results from Table 2 for the MOM 12/12 portfolio illustrates the point that a simple
t-ratio test on the mean is not equivalent to a statistical arbitrage test. The MOM 12/12
portfolio registers a t-statistic of 3 on the mean and yet fails to be a statistical arbitrage.
This is because its time-averaged variance is not decreasing, as suggested in Figure 2. This
observation is more apparent when we compare the cumulative prots from the MOM 12/12
portfolio to those of the MOM 6/9 portfolio (which tests positively for statistical arbitrage).
For comparative purposes with Figure 1, we use Monte Carlo simulation (once again with
10,000 trials) to generate an empirical distribution for the cumulative trading prots of the
MOM 6/9 trading strategy. The empirical distributions of the payos are plotted in Figure
3 over the same four horizons and are consistent with the underlying normal distribution.
The results for the MOM 6/9 strategy compare favorably to those of Example 1. Observe
that the mean payos in Figure 3 are higher than the corresponding mean payos in Figure
1 while the variances in Figure 3 are smaller. Indeed, the time-averaged variance of MOM
6/9 is declining, in contrast to Example 1. In addition, the probability of a loss declines
much more rapidly in Figure 3 than Figure 1.
To get a sense of how fast the statistical arbitrages are converging to riskless arbitrages,
we plot the probability of a loss, using equation (9), for two representative momentum
strategies that constitute statistical arbitrages at the 5% level: MOM 6/12 and MOM 9/12
(see Figures 4 and 5). The time-averaged variances of the strategies are plotted as well.
For the MOM 6/12 strategy, after just 89 months, the probability of incurring a loss on
this strategy falls below 1%. A casual observation of the time-averaged variance graphs in
Figures 4 and 5 reveals that the MOM 6/12 converges more rapidly to a riskless arbitrage
than MOM 9/12. This observation is corroborated by the p-values and estimates reported
in Table 2.
22
5.2 Value strategies
Lakonishok, Shleifer, and Vishny (1994) are among the rst to examine a comprehensive set
of contrarian strategies. They focus on value strategies constructed from stock ratios such as
book-to-market, cash ow to price, and earnings to price and on past performance variables
like sales growth. They nd evidence to suggest that such value strategies are protable
for holding periods ranging from one year to ve years. What remains an open question is
whether these value strategies contradict the concept of market eciency.
In this section, we test whether the strategies examined by Lakonishok, Shleifer, and
Vishny (1994) constitute statistical arbitrage opportunities. Specically we test strategies
that long the top decile and short the bottom decile of stocks based on book-to-market,
cash ow-to-price, or earnings-to-price ratios of the past year. We hold these portfolios for
one, three or ve years. We also test a value strategy based on the past three years of sales
growth. We long the bottom decile and short the top decile of stocks based on the past three
years of sales growth and hold this spread for one, three or ve years.
The results in Table 3 suggest that value strategies based on past three years of sales
growth are consistent with statistical arbitrage. For all holding periods considered, the sales
growth based value strategy tests positively for statistical arbitrage at the 5% level. Value
strategies based on cash ow to price also exhibit this tendency. For holding periods of
three and ve years, the cash ow-to-price based value strategies yield statistical arbitrages
at the 5% level. However, their expected prots are lower than those of their sales growth
based counterparts. The value strategies based on the other variables do not yield statistical
arbitrages at the 5% level. In particular, the point estimates for the variance growth rates
of earnings-to-price based strategies (holding period equals one and three years) are positive
(though not signicantly so) indicating that these strategies may have become riskier over
time.
Figures 6 and 7 plot the probability of loss and time-averaged variance for two repre-
sentative value strategies which constitute statistical arbitrage at the 5% level. Supporting
the results of Table 2, both the probability of loss and the time-averaged variance of the
sales-based value strategy with a holding period of ve years are lower than those of the
cash ow to price-based value strategy with a holding period of ve years.
23
Consistent with the results from Fama and French (1992), among others, the expected
prots of strategies based on the book-to-market ratio are statistically greater than zero both
from the t-ratio test results and from the H1 test (which is joint with H2) results. However,
the growth rate of the variance for such strategies is not statistically less than zero, at the 5%
level. Therefore, book-to-market based value strategies do not constitute statistical arbitrage
opportunities. This dovetails with the Fama and French (1993) notion that high book-to-
market stocks may be systematically riskier than low book-to-market stocks and therefore
are entitled to higher expected returns.
The reason why the cash ow-to-price and earnings-to-price strategies do not perform as
well as they do in Lakonishok, Shleifer, and Vishny (1994) may be due to the dierence in
the sample period. Lakonishok, Shleifer, and Vishny (1994)s sample period ends in 1989
while our sample period extends till 2000. In results not shown, plots of the cumulative
discounted trading prots generated by cash ow-to-price and earnings-to-price based value
strategies suggest that these strategies do not perform as well in the post-1990 period as in
the pre-1990 period.
Overall, roughly half the value strategies considered (ve out of 12) provide evidence
against market eciency. This result is similar to that from testing momentum trading
strategies. For illustration, we bootstrap the residuals from the estimation of mean incre-
mental prot and the growth rate of the variance. The bootstrap samples of the mean
incremental prot can be
6 Robustness checks
In this section we perform a series of tests to gauge the robustness of the results from the
previous section. First, we evaluate the appropriateness of the constrained mean (CM)
model. Second, we test the sensitivity of the statistical arbitrage tests to deviations from
our assumed incremental trading process. Third, we measure the impact of market frictions
like transactions costs, margin requirements, liquidity buers, and a higher borrowing rate
on the incremental prots for the momentum and value strategies, and retest for statistical
arbitrage. Fourth, we check for the inuence of small stocks on our results.
24
6.1 Testing the constrained mean model
The rst order of business is to verify whether the constrained mean (CM) model, which
assumes expected trading prots are constant over time, oers a good t for the incremental
prot processes we examine. To do so, we compare measures of t for the CM model
with those of the unconstrained mean (UM) model. We also study the estimated rates of
change in the expected trading prots to determine whether they are signicantly negative
and implement a likelihood ratio test. From a statistical perspective, introducing the
parameter, whose point estimates are usually negative, reduces the number of signicantly
positive estimates. Overall, for the UM test, the point estimates of increase along with
their standard errors as the information contained in trading prots is spread over a fourth
parameter.
The rst measure of t we examine is the average root mean squared error (RMSE). The
RMSE numbers are calculated by comparing the portfolios observed incremental trading
prots with those from 10,000 simulations. Each simulated time series is based on para-
meters estimated from the observed incremental trading prots of a given trading strategy.
Specically, we simulate an identical number of incremental trading prots as the observed
sample, using each trading strategys parameter estimates. We then compute the RMSE
between the observed trading prots and those from the simulation and repeat 10,000 times
with the average RMSE reported in Table 4. If the UM model oers a better t of the
data than the CM model, then the RMSE numbers for the former should be lower than
those of the latter. However, Table 4, which records the RMSE numbers for both models,
demonstrates that the UM model does not oer a better statistical alternative than its CM
counterpart. In fact, the RMSE numbers are almost indistinguishable.
The second measure of t we look at is the sum of normalized squared residuals (SSR).
If the SSR numbers of the UM model are lower than those of the CM model, then we are
lead to conclude that it is the better model for the data. Yet Table 4 shows that the SSR
numbers for the two models are once again almost identical.
Next, we test whether the incremental prots of the portfolios are decreasing using a
t-statistic. The p-values for this test are also recorded in Table 4. We nd that for all except
25
one
15
of the portfolios which are CM statistical arbitrages at the 10% or 5% level, the growth
rate of the incremental prots is statistically indistinguishable from zero. Therefore, there
is no need to estimate and weaken the power of the statistical arbitrage test.
More formally, we employ a likelihood ratio test
16
for the restriction = 0 and com-
pare the results with a
2
1,0.10
test statistic equal to 2.71. For all trading strategies under
consideration, the null hypothesis that = 0 is accepted without reservation.
In summary, the UM test for statistical arbitrage unnecessarily increases the standard
errors of the mean incremental prot estimates without oering a better goodness of t as
compensation. Hence the CM model is more appropriate for modeling observed incremental
trading prots.
6.2 Accuracy with autocorrelation, jumps, and non stationarity
To gauge the robustness of our assumed process in equation (1), simulations are conducted to
investigate the impact of autocorrelation, jumps, and non stationary parameters on inferences
regarding the presence of statistical arbitrage. Using parameters , , and consistent
with no statistical arbitrage
17
and statistical arbitrage, a time series of 400 incremental
trading prots are simulated, corresponding to the approximate number of months in our
sample period. The parameter used to generate autocorrelation following the MA(1)
process specied in Assumption 2 of Appendix A.4 is 0.40. This parameter corresponds
to a correlation between incremental trading prots of 0.34

1+
2
as seen in Appendix
15
Interestingly, this portfolio (MOM 9/12) is a statistical arbitrage according to the UM test. Therefore,
for the one instance where trading prots are better modeled using a UM process with a signicantly negative
estimate, the same conclusion is reached regarding market eciency.
16
For the likelihood ratio test, abbreviated LRT,
2 ln(LRT) = 2
n

i=1
ln

R
i

+
n

i=1
(v
i

R
)
2

2
R
i
2R
2
n

i=1
ln

UR
i

UR

+
n

i=1

v
i

UR
i
UR

2
UR
i
2UR
and this statistic must be greater than a
2
variable with one degree of freedom at a given level in order
to reject the null hypothesis that = 0. The subscripts R and UR denote the restricted and unrestricted
parameter estimates resulting from two separate estimations. See Theorem 8.4.1 Casella and Berger (1990)
page 380.
17
No null hypotheses for the autocorrelation, jump, or Markov chain parameters are required as we are
only interested in studying their impact on results derived under Assumption 1.
26
A.4, and is similar to the average estimate of our trading strategies. Jumps are simulated
from a normal distribution while non stationary parameters are modeled using a two state
Markov chain with low and high values for and . Statistical arbitrage opportunities are
simulated using parameters = 0.01, = 0.02, and = 0.15 while trading prots that are
not consistent with statistical arbitrage are generated by the same , and parameters but
= 0. Dierent values for are considered with the statistical procedure performing perfectly
whenever > 0, with 0 producing similar results. Therefore, = 0 ensures the test for
statistical arbitrage is as dicult as possible and this value is used to dierentiate between
an economy with and without the existence of statistical arbitrage. For completeness, many
other possible parameter values are investigated with little change to the nal results.
As a second step, estimation involving Lemma 5, based on the assumed process in equa-
tion (1), is conducted on each simulated time series. A decision regarding statistical arbitrage
is rendered according to Theorem 6. A total of 10,000 time series are simulated in order to
heuristically ascertain the power of our statistical test.
The intensity parameter for jumps yields an average of 1 jump every 3 years with jump
magnitudes normally distributed with zero mean and standard deviation equal to 0.05. A
two state Markov chain generates non stationary drift and volatility parameters according
to the transition matrix

0.8 0.2
0.2 0.8

for low and high and values. The low and high values for are 0.005 and 0.01, respec-
tively, while had low and high values of 0.015 and 0.03. The parameter values for are
once again 0.15 and 0 to simulate circumstances with and without statistical arbitrage,
respectively.
The number of conclusions that are altered by autocorrelation, jumps, or non stationary
parameters can be inferred from Table 5. Observe that the statistical test correctly accepts
the null hypothesis of no statistical arbitrage with great accuracy, even when there exists a
deviation from the assumed process. At the 10% level, when the underlying parameters are
not consistent with statistical arbitrage, the test rejects the null hypothesis no more than
5% of the time.
27
Overall, the performance of the statistical test is exceptional even with deviations in
the assumed process. Clearly, the presence of autocorrelation, jumps, and parameter non
stationarity lead to a bias towards accepting the null hypothesis of no statistical arbitrage.
Hence, the formulation in equation (1) may be considered fail-safe in the presence of
deviations. This property stems from the fourth property of statistical arbitrage. The
additional volatility caused by jumps and non stationary parameters increases the standard
error of , which translates into higher corresponding p-values for H2, and consequently a
higher probability of accepting the null hypothesis of no statistical arbitrage. Along with the
Bonferroni inequality, these simulations demonstrate that our test for statistical arbitrage is
very conservative when rejecting the null hypothesis of no statistical arbitrage, allowing one
to condently reject market eciency.
6.3 Transactions costs and market frictions
A common critique of nancial anomalies is that the trading prots from such anomalies
disappear after adjusting for transactions costs. Therefore, we estimate the transactions
costs of the portfolios in Section 4 as follows. First, we estimate the average monthly
turnover for each of the portfolios by taking a ratio of the sum of buys and sells each period
over two times the total number of stocks held in that period. This estimates the number
of round trip transactions as a percentage of the number of stocks held. Then we multiply
this measure of monthly turnover with an estimate of the round trip transactions cost given
in Chan and Lakonishok (1997). According to Chan and Lakonishok (1997), the average
stock in the NASDAQ has a round trip transactions cost
18
of 1.34%. Finally, we adjust the
monthly prots downward by the transactions costs.
We also consider the impact of four market frictions on the ability of momentum and
value trading strategies to yield statistical arbitrage. Our analysis is based on the exhaustive
approach of Alexander (2000). The rst three frictions consist of the margin (m) required
for both long and short positions, the additional margin known as a liquidity buer for the
marking-to-market of short positions (c), and the haircut brokerages implicitly levy on their
18
We also calculate transactions costs using the higher NYSE estimate of 1.55% and re-test for statistical
arbitrage. None of the inferences change with the NYSE estimate for round trip transactions cost.
28
clients by withholding a small percentage of the interest earned on their margin accounts
(h). Moreover, we also impose a higher borrowing rate for nancing trading losses than the
lending rate for investing trading prots.
More specically, the adjusted return on the portfolio denoted r
adj
equals
r
adj
= (r
unadj
h)/(2m+c) (10)
as in equation (3) of Alexander (2000), where r
unadj
represents the unadjusted portfolio
return used in our previous analysis. See Alexander (2000) for a detailed discussion.
Jacobs and Levy (1995) estimate the haircut on the margin account to be between 25
and 30 basis points per year while Jacobs and Levy (1997) nd the liquidity buer to be
10% of short sale proceeds. Hence, we select a liquidity buer for short sales of 10%, and a
haircut of 27.5 basis points. Furthermore, we set a conservative
19
margin rate of 50%. The
borrowing rate is chosen to be 2% higher (per year) than the lending rate. Using AA 30
day commercial paper for nancial rms available from the St. Louis Federal Reserve, we
nd the maximum spread over Treasury Bills of the same maturity is 1.7% from 1997 to
2002, the period for which data is available. Therefore, our choice of a 2% spread is very
conservative. The higher borrowing rate only applies in those infrequent instances when the
cumulative trading prot for a statistical arbitrage is negative. Thus, a higher lending rate
is unlikely to have signicant repercussions for our previous analysis.
After incorporating these four frictions and accounting for transaction costs, the results
regarding the existence of statistical arbitrage in Table 6 are nearly identical to previous
results in Tables 2 and 3. Although the estimated parameters are smaller, of the 6 mo-
mentum trading strategies that originally test positive for statistical arbitrage at the 5%
level, 5 remain statistical arbitrages at this level. For value trading strategies, 4 of the 5
strategies remain statistical arbitrages at the 5% level while the other represents a statistical
arbitrage at the 10% level. Thus, the ability of momentum and value trading strategies to
yield statistical arbitrage does not disappear in the presence of realistic trading frictions.
19
See the Securities and Exchange Commission website http://www.sec.gov/investor/pubs/margin.htm
for details regarding Regulation T and justication for our choice of 50%.
29
6.4 Inuence of small stocks
Several tests of stock market anomalies conclude that small stocks are more inecient than
large stocks (Hong, Lim, and Stein, 2000; Mitchell and Staord, 2000). In addition, it is
well-known that small stocks are less liquid, and have greater transactions costs than large
stocks (Chan and Lakonishok, 1997). Therefore, it may be useful to see if our results survive
the exclusion of these illiquid stocks.
To this end, we remove stocks in our sample with market capitalizations below the 50th
NYSE market equity percentile (which is about 1 billion dollars in 2000). This reduces our
sample size by more than two-thirds since there are many more small stocks than large stocks
and there are relatively more small stocks traded in the NASDAQ (which are included in
the CRSP database) than in the NYSE. We then re-test our trading strategies for statistical
arbitrage.
Results from the tests of statistical arbitrage suggest that evidence of statistical arbitrage
is even greater without the bottom 50th NYSE percentile. All but one of the momentum
strategies and 7 out of 12 value strategies are statistical arbitrages at the 5% level. Further,
all the momentum portfolios which are statistical arbitrages for the full sample are also
statistical arbitrages for the sample less the bottom 50th NYSE market equity percentile.
The value portfolios are less robust to the variation in sample composition. The sales based
strategies fail to be statistical arbitrages with the reduced sample. However there exist
strategies based on cash ow-to-price, book-to-market equity and earnings-to-price which
become statistical arbitrages following the removal of small stocks. This clearly indicates
that our results are not driven by small stocks.
6.5 Sensitivity to the sequence of returns
In response to concerns that statistical arbitrage may be an artifact of a particular observed
sequence of returns, a further robustness check is conducted on the and parameters of
all trading strategies. This additional robustness check involves bootstrapping t-statistics
to ascertain their distribution, and re-arms our conclusions regarding statistical arbitrage.
Our emphasis is on reporting the results for , the parameter most often responsible for
determining whether a trading strategy constitutes statistical arbitrage. In most cases,
30
when is not signicantly positive, is not signicantly negative (with possible exceptions
being CP3 and CP5 whose results are reported below).
We implemented the following test procedure to verify whether the estimates for (as
well as ) are indeed statistically negative (positive):
1. Compute residuals based on the sample estimates of , , and . This sequence
of residuals is the basis for the bootstrap procedure in the next step which alters the
sequence of residuals, hence returns, according to a set of uniformly distributed integers
based on the sample size of each individual strategy.
2. Bootstrap 10,000 sequences of trading prots using the residuals from the original
sequence under the assumption that = 0 while and are set equal to their sample
estimates.
3. For each of the 10,000 trials, compute the corresponding t-statistic for . Since we are
testing whether is negative, then nd the 5
th
percentile of these simulated t-statistics
as smaller values are indicative of greater signicance. Denote the 5
th
percentile of the
bootstrapped t-statistics as the critical t-statistic, abbreviated t
crit
.
4. Compare the sample t-statistic with t
crit
for each trading strategy. If the sample t-
statistic is lower than t
crit
, reject the hypothesis that is non-negative and conclude
the trading strategy is indeed a statistical arbitrage.
We nd the values of t
crit
are similar for the various trading strategies, with most larger
than -1.70. This result supports our earlier conclusions reported in Tables 2 and 3. For
example, the MOM 9/12 strategy has a t
crit
of -1.655 while its sample t-statistic is -3.031,
allowing us to condently reject the null hypothesis and conclude that is negative, as
required for statistical arbitrage. Similarly, for the SALES 5 strategy, t
crit
equals -1.641
while the sample t-statistic is -4.176.
Furthermore, we also conrm that our rejections of statistical arbitrage are justied. For
example, the MOM 12/12 strategy has a critical value of -1.652, smaller than the sample
t-statistic of 0.012. This outcome is consistent with accepting the null hypothesis that is
non-negative, reinforcing the conclusion that MOM 12/12 is not a statistical arbitrage.
31
When testing whether is statistically positive, we are interested in the 95
th
percentile
of the bootstrapped t-statistics to determine t
crit
. For CP3 and CP5, the t
crit
statistics for
from the bootstrap procedure are 1.781 and 1.713 respectively, both of which are below
the sample t-statistics. Thus, we are able to condently reject the hypothesis that is
non-positive and conclude that CP3 and CP5 remain statistical arbitrage opportunities.
In summary, the bootstrap procedure in this subsection conrms that our decisions re-
garding a trading strategys ability to produce (or fail to produce) statistical arbitrage is not
unduly inuenced by the observed sequence of returns.
6.6 Final comments on robustness
It is important to emphasize the conservative nature of our statistical test. First, despite
the accuracy of our estimation procedure, the Bonferroni inequality requires the sum of the
p-values to be below the level. Second, deviations from our assumed process bias the test
against rejecting the null hypothesis of no statistical arbitrage. Hence, evidence of statistical
arbitrage is unlikely the consequence of a misspecied trading prot process.
For completeness and potential applications to other trading strategies in future research,
a brief summary of statistical tests that account for autocorrelated trading prots is pro-
vided in Appendix A.4. Empirically implementing this material does not lead to dierent
conclusions regarding market eciency for the trading strategies under consideration in this
paper. Furthermore, we consider the maximum loss incurred by pursuing our statistical
arbitrage opportunities which equals the minimum value of the cumulative trading prots.
These values are reported in Table 6 and proxy for the amount of capital an investor requires
to engage in trading strategies that oer statistical arbitrage. These values are not seen to
be prohibitive, especially in light of the potential rewards.
7 An application: The size eect
The size eect of Banz (1981) oers an interesting test case to apply our proposed statistical
arbitrage methodology as it is often believed to have disappeared after the mid 1980s. Thus,
it may be interesting to investigate whether the strategy of longing small stocks and shorting
32
large stocks constitutes a statistical arbitrage. Using the Fama-French SMB factor as a
proxy for the size eect, we consider the incremental and cumulative trading prots based
on a trading strategy which longs stocks with market capitalizations below the 30th NYSE
percentile and shorts stocks with market capitalizations above the 70th NYSE percentile.
The simple t-statistic on the mean prot is 2.11, consistent with positive expected prots.
However this does not imply that the SMB based trading strategy violates market eciency.
Indeed, applying the statistical arbitrage methodology yields parameter estimates for ,
, and of 0.0002, 0.6404, and -0.7961 respectively, while the p-value for statistical arbitrage
equals 0.294. Hence, the size eect does not constitute a statistical arbitrage. A plot of the
cumulative prots from the SMB based strategy in Figure 8 reveals that the SMB trading
prots dissipate around the mid 1980s. In fact, its mean incremental prot after 1985 is
negative, implying that statistical arbitrage portrays the market eciency implications of
the size eect more accurately than a t-ratio test. In summary, our ndings corroborate the
popular belief that the size eect disappeared around the mid 1980s.
8 Conclusion
This paper introduces the concept of statistical arbitrage which facilitates a test of market
eciency without the need to specify an equilibrium model. In the limit, statistical arbitrage
converges to arbitrage. Consequently, the joint hypothesis dilemma is avoided by appealing
to a long horizon trading strategy. Methodologies that test for the existence of statistical
arbitrage are provided. Simulations reveal the robustness and the conservative nature of our
statistical test, with a bias towards accepting the null hypothesis of no statistical arbitrage in
the presence of deviations from our assumed trading prot process such as autocorrelation,
jumps, and non stationary parameters.
Tests for statistical arbitrage are applied to the cumulative discounted trading prots of
momentum (Jegadeesh and Titman, 1993) and value trading strategies (Lakonishok, Shleifer,
and Vishny, 1994). In contrast to the t-test of positive risk-adjusted expected returns, the
existence of statistical arbitrage is sucient to reject market eciency for any model of
market returns.
Roughly half of the momentum and value strategies we evaluate test positively for statis-
33
tical arbitrage. This suggests the existence of several trading opportunities that converge to
riskless arbitrages with decreasing time-averaged variances, a result dicult to reconcile with
market eciency. Our test results are robust to the exclusion of small stocks and to the in-
corporation of several market frictions like transactions costs, margin accounts, short-selling
buers, and higher borrowing rates.
Promising avenues of research include testing other anomalies for statistical arbitrage,
such as the abnormal returns from dividends or earnings announcements. Also, the Bonfer-
roni inequality may be replaced by a more computationally intensive monte carlo procedure
in order to simulate the critical values underlying our joint hypothesis test.
34
References
[1] Alexander, G.J., 2000. On Back-Testing Zero-Investment Strategies. Journal of Busi-
ness 73, 255-278.
[2] Banz, R. W., 1981. The relationship between return and market value of common stocks.
Journal of Financial Economics 9, 3-18
[3] Bernardo, A.E., O. Ledoit, 2000. Gain, loss, and asset pricing. Journal of the Political
Economy 108, 144-172.
[4] Bodurtha, Jr. J.N., G.R. Courtadon, 1986. Eciency tests of the foreign currency op-
tions market. Journal of Finance 41, 151-162.
[5] Box, G.E.P., G.M. Jenkins, 1976. Time series analysis: Forecasting and control. Holden-
Day.
[6] Carhart, M., 1997. On persistence of mutual fund performance. Journal of Finance 52,
57-82.
[7] Carr, P., H. Geman, D. Madan, 2001. Pricing and hedging in incomplete markets.
Journal of Financial Economics 62, 131-167.
[8] Casella, G., R.L. Berger, 1990. Statistical inference. Duxbury Press.
[9] Chan, L.K.C., N. Jegadeesh, J. Lakonishok, 1996. Momentum strategies. Journal of
Finance 51, 1681-1713.
[10] Chan, L.K.C., J. Lakonishok, 1997. Institutional equity trading costs: NYSE versus
NASDAQ. Journal of Finance 52, 713-735.
[11] Cochrane, J.H., J. Sa a-Requejo, 2000. Beyond arbitrage: Good-deal asset price bounds
in incomplete markets. Journal of the Political Economy 108, 79-119.
[12] Due, D., 2001. Dynamic asset pricing theory, third edition. Princeton University Press.
[13] Dybvig, P.H., C. Huang, 1988. Nonnegative wealth, absence of arbitrage, and feasible
consumption plans. Review of Financial Studies 1, 377-401.
35
[14] Fama, E.F., K. R. French, 1992. The cross-section of expected stock returns. Journal of
Finance 47, 427-465.
[15] Fama, E.F., K.R. French, 1993. Common risk factors in the returns of stocks and bonds.
Journal of Financial Economics 33, 3-56.
[16] Fama, E.F., 1998. Market eciency, long-term returns, and behavioral nance. Journal
of Financial Economics 49, 283-306.
[17] Hong, H., T. Lim, J. Stein, 2000. Bad news travels slowly: Size, analyst coverage, and
the protability of momentum strategies. Journal of Finance 55, 265-295.
[18] Hull, J.C., 1997. Options, futures, and other derivatives, third edition. Prentice-Hall.
[19] Jacobs, B.I., K.L. Levy, 1995. More on long-short strategies. Financial Analysts Journal
51, 88-90.
[20] Jacobs, B.I., K.L. Levy, 1997. The long and short on long-short. Journal of Investing 6,
73-86.
[21] Jarrow, R., 1988. Finance theory. Prentice-Hall.
[22] Jegadeesh, N., S. Titman, 1993. Returns to buying winners and selling losers: Implica-
tions for stock market eciency. Journal of Finance 48, 65-91.
[23] Jensen, M.C., 1978. Some anomalous evidence regarding market eciency. Journal of
Financial Economics 6, 95-101.
[24] Kamara, A., T.W. Miller, Jr., 1995. Daily and intradaily tests of european put-call
parity. Journal of Financial and Quantitative Analysis 30, 519-539.
[25] Lakonishok, J., A. Shleifer, R.W. Vishny, 1994. Contrarian investment, extrapolation,
and risk. Journal of Finance 49, 1541-1578.
[26] LeRoy, S.F., 2002. Innite portfolio strategies. Working paper, University of California,
Santa Barbara.
36
[27] Mitchell, M., E. Staord, 2000. Managerial Decisions and long-term stock price perfor-
mance. Journal of Business 73, 287-329.
[28] Ross, S., 1976. The arbitrage theory of capital asset pricing. Journal of Economic Theory
13, 341-360.
[29] Shleifer, A., R.W. Vishny, 1997. The limits of arbitrage. Journal of Finance 52, 35-55.
[30] Shleifer, A., 2000. Inecient markets: An introduction to behavioral nance. Oxford
University Press.
37
A Appendices
A.1 Importance of fourth condition
Consider the standard Black-Scholes economy with
S
t
= S
0
e
t
2
t/2+Wt
and assume r = 0 implying B
t
1 for all t. This economy is in equilibrium but would be
rejected by a denition of statistical arbitrage that does not include the fourth condition.
To prove this, rst consider a simpler economy with stock dynamics
s
t
= s
0
+t
2
t/2 +W
t
and a buy and hold strategy similar to Example 2 whose value at time t equals v(t) =
1 s
t
s
0
1. For r = 0, cumulative trading prots are identical to discounted cumulative
trading prots. Cumulative trading prot dynamics are represented as
v(t) = s
0
+t
2
t/2 +W
t
s
0
= [
2
/2]t +W
t
.
If
2
/2 > 0, then P( v(t) < 0) converges to zero. This claim follows from Appendix
A.3 with and both equaling zero and =
2
/2. This portfolio would also satisfy
conditions 1 and 2 of statistical arbitrage. Returning to the Black-Scholes economy with
S
0
= 1, consider the same buy and hold trading strategy with cumulative trading prots
v(t) = 1 S
t
S
0
1 = S
t
1 = e
st
1.
The following sequence illustrates that P(v(t) < 0) also converges to zero
P(v(t) < 0) = P(e
st
1 < 0) < P( s
t
1 < 0) = P( v(t) < 0)
where the second relation employs the inequality S
t
= e
st
> s
t
. Thus, without condition 4,
the denition of statistical arbitrage would apply to a buy and hold strategy in the Black-
Scholes economy which is known to be in equilibrium.
38
A.2 Doubling strategies
The canonical discrete time doubling strategy is a fair coin toss with an investor winning
$1 if the coin toss is heads, and doubling their bet if the coin toss is tails. This strategy
continues until the rst heads is attained, at which time the strategy is stopped. LeRoy
(2002) also considers such a strategy (equation (15) of Section 5). To demonstrate
20
that
our empirical ndings are not the result of a doubling strategy, we consider their cumulative
trading prots v(t
i
) which equal 1, 3, 7, 15, . . . and so on for t
i
= 1, 2, 3, 4, . . . until
the stopping time t

when v(t

) = 1. Thus, an investor collects one dollar, after accounting

for the series of increasingly large previous losses, when the rst heads is attained.
Until the stopping time is reached, all incremental trading prots are negative with
the investor losing 2
t
i
1
dollars at t
i
= 1, 2, 3, 4, . . . before t

. The doubling strategy has

cumulative trading prots equal to
v(t
i
) =

1 with probability 1

1
2

t
i
2
t
i
+ 1 with probability

1
2

t
i
and converges almost surely to a riskless arbitrage although v(t
0
= 0) = 0. However, the
doubling strategy is not a statistical arbitrage opportunity since the expectation and variance
of v(t
i
) are zero (rather than positive) and 2
t
i
1 respectively. Therefore, the time-averaged
variance does not decline as
2
t
i 1
t
i
. In summary, doubling strategies violate the second
and fourth properties of statistical arbitrage in Denition 4.
A.3 Zero probability of loss in limit
Equation (2) states that v(t
n
) is distributed
N

i=1
i

,
2
n

i=1
i
2

.
Convergence of the tail probability, P(v(t
n
) < 0), to zero is determined by the cumulative
normal distribution
P (v(t
n
) < 0) =
1

2
2

n
i=1
i
2

1
2
2
!
n
i=1
i
2
(x
!
n
i=1
i

)
2
dx.
20
Without loss of generality, we assume the riskfree interest rate is zero.
39
The above probability may be expressed as
P (v(t
n
) < 0) = P

Z <

n
i=1
i

n
i=1
i
2

= N

n
i=1
i

n
i=1
i
2

where Z is a standard normal and N(x) is the standard normal cumulative distribution
function. Although there is no closed form solution for the standard normal cdf, a polynomial
approximation (for x < 0) is available, Hull (1997) page 243
N(x) = N

(x)

a
1
1
1 +x
+a
2
1
(1 +x)
2
+a
3
1
(1 +x)
3
+ h.o.t.

where a
1
= 0.4361836, a
2
= 0.1201676, a
3
= 0.9372980 and = 0.33267 are constants.
The leading term governs the rate of convergence implying
P (v(t
n
) < 0) = O

a
1

2
exp

n
i=1
i

2
2
2

n
i=1
i
2

1
1 +

!
n
i=1
i

!
n
i=1
i
2

which converges to zero as n increases under the assumption of statistical arbitrage. In

particular, the probability converges faster than rate
O

exp

n
i=1
i

2
2
2

n
i=1
i
2

as a function of n.
A.4 Correlated trading prot increments
Trading strategies may produce autocorrelated trading prots. For example, the momen-
tum trading strategies have prots derived from overlapping periods. Consider an MA(1)
process to instill correlation between consecutive
21
trading prot increments according to
the parameter . Specically, the correlation between v
i
and its previous value v
i1
is
approximately

1+
2
under the following assumption.
Assumption 2 Let the discounted incremental trading prots satisfy
v
i
= i

+i

z
i
+(i 1)

z
i1
for i = 1, 2, . . . , n where z
i
are i.i.d N(0,
2
) random variables with z
0
= 0.
21
Generalizing the process to an MA(q) for q 2 follows in a straightforward manner.
40
For parameter estimation, several iterative procedures exist as detailed in Box and Jenk-
ins (1976) chapter 7. The distribution of v(t
n
) is required to construct a test for statistical
arbitrage and determine the convergence rate to arbitrage. The distribution of the cumula-
tive discounted trading prots is the topic of the next lemma.
Lemma 8 The distribution of v(t
n
) with correlated increments equals
v(t
n
) = N

i=1
i

,
2

(1 +)
2
n1

i=1
i
2
+n
2

. (11)
Proof. The distribution of v(t
n
) is solved by recursion
v(t
i
) = v(t
i1
) + i

+i

z
i
+(i 1)

z
i1
= v(t
i2
) + (i 1)

+ (i 1)

z
i1
+(i 2)

z
i2
+ i

+i

z
i
+(i 1)

z
i1
= v(t
i3
) + (i 2)

+ (i 2)

z
i2
+(i 3)

z
i3
+ (i 1)

+ (i 1)

z
i1
+(i 2)

z
i2
+ i

+i

z
i
+(i 1)

z
i1
and so forth generating the following solution for v(t
n
)
v(t
n
) =
n

i=1
i

+ (1 +)
n1

i=1
i

z
i
+n

z
n
= N

i=1
i

,
2

(1 +)
2
n1

i=1
i
2
+n
2

. (12)
As expected, equation (11) reduces to equation (2) when = 0.
Theorem 9 Under the previous assumption, a trading strategy generates a statistical arbi-
trage with 1 percent condence if the following conditions are satised:
H1: > 0 ,
H2:

< 0 ,
H3:

> max

1
2
, 1

,
with the sum of the individual p-values forming an upper bound for the Type I error.
41
Proof. The variance in equation (12) may be modied to equal
2
(1 + )
2

n
i=1
i
2
by
adding the term
2
(2 +
2
)n
2
. The marginal contribution of this additional term

(1 +)
2

n1
i=1
i
2
+n
2

+
2
(2 +
2
)n
2

(1 +)
2

n1
i=1
i
2
+n
2
0
as n increases, ensuring asymptotic inferences remain valid. Consequently, for large n, the
approximation v(t
n
)
d
N

n
i=1
i

,
2
(1 +)
2

n
i=1
i
2

is analyzed. Therefore, testing

for statistical arbitrage involves the same exact joint hypotheses as Theorem 6 but with a
modied variance that accounts for .
The next theorem follows immediately from Theorem 7.
Theorem 10 The rate of convergence from statistical arbitrage to arbitrage is exponential
with respect to n. Specically, the convergence rate is faster than
O

exp

n
i=1
i

2
2
2
(1 +)
2

n
i=1
i
2

.
Theorems 9 and 10 illustrate that, conditional on parameters ,
2
, , and being equal,
positive correlation implies slower convergence (higher variance) with (1 + )
2
> 1 while
negative correlation implies faster convergence (lower variance) with (1 +)
2
< 1.
42
Fig 1. Discounted cumulative trading profits from a theoretical trading strategy which invests in a risky asset whose price follows a
geometric Brownian motion and borrows at the risk-free rate (Example 1 in the text). 10000 Monte Carlo trials are used to obtain the
distribution of the payoffs from the above theoretical trading strategy. The frequency distribution of the cumulative payoffs for one,
two, three, and five years are illustrated.
-2 -1 0 1 2 3 4 5
0
200
400
600
800
1000
One Year
Payoff
F
r
e
q
u
e
n
c
y
-2 -1 0 1 2 3 4 5
0
200
400
600
800
1000
Two Years
Payoff
F
r
e
q
u
e
n
c
y
-2 -1 0 1 2 3 4 5
0
200
400
600
800
1000
Three Years
Payoff
F
r
e
q
u
e
n
c
y
-2 -1 0 1 2 3 4 5
0
200
400
600
800
1000
Five Years
Payoff
F
r
e
q
u
e
n
c
y
Fig 2. Comparison of cumulative trading profits from two momentum strategies. Sample period is from January 1965 to December 2000. The
momentum strategy MOM 6/9 longs the top return decile of stocks and shorts the bottom return decile of stocks in the sample based on a
formation period of 6 months and a holding period of 9 months. The momentum strategy MOM 12/12 longs the top return decile of stocks and
shorts the bottom return decile of stocks in the sample based on a formation period of 12 months and a holding period of 12 months. $1 is
invested at all times in both portfolios. Any profits (losses) are re-invested in (financed with) the risk free asset.
Jan 1970 Jan 1980 Jan 1990 Jan 2000
-0.5
0
0.5
1
1.5
2
2.5
Time (months)
C
u
m
u
l
a
t
i
v
e

p
r
o
f
i
t

i
n

d
o
l
l
a
r
s
Cumulative Trading Profits for MOM 6/9 Portfolio
Jan 1970 Jan 1980 Jan 1990 Jan 2000
-0.5
0
0.5
1
1.5
2
2.5
Time (months)
C
u
m
u
l
a
t
i
v
e

p
r
o
f
i
t

i
n

d
o
l
l
a
r
s
Cumulative Trading Profits for MOM 12/12 Portfolio
Fig 3. Discounted cumulative trading profits from a momentum (MOM 6/9) trading strategy. The momentum strategy MOM 6/9
longs the top return decile of stocks and shorts the bottom return decile of stocks in the sample based on a formation period of 6
months and a holding period of 9 months. $1 is invested at all times. Any profits (losses) are re-invested in (financed with) the risk
free asset. 10000 Monte Carlo trials are used to obtain the empirical distribution of the payoffs from the above trading strategy. The
frequency distribution of the cumulative payoffs for one, two, three, and five years are illustrated.
Fig 4. Probability of a loss and time-averaged variance for the momentum strategy (MOM 6/12) which represents a CM statistical
arbitrage opportunity. Sample period is from January 1965 to December 2000. The momentum strategy (MOM 6/12) longs the top
return decile of stocks and shorts the bottom return decile of stocks in the sample based on a formation period of 6 months and a
holding period of 12 months. $1 is invested at all times. Any profits (losses) are re-invested in (financed with) the risk free asset. The
probability of loss is calculated as per equation (9).
0 20 40 60 80 100 120
2
4
6
8
10
x 10
-4
Months
T
i
m
e
-
a
v
e
r
a
g
e
d

v
a
r
i
a
n
c
e
0 20 40 60 80 100 120
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
Months
P
r
o
b
a
b
i
l
i
t
y

o
f

L
o
s
s
Probability of loss
Time-averaged variance
Fig 5. Probability of a loss and time-averaged variance for the momentum strategy (MOM 9/12) which represents a CM statistical
arbitrage opportunity. Sample period is from January 1965 to December 2000. The momentum strategy (MOM 9/12) longs the top
return decile of stocks and shorts the bottom return decile of stocks in the sample based on a formation period of 9 months and a
holding period of 12 months. $1 is invested at all times. Any profits (losses) are re-invested in (financed with) the risk free asset. The
probability of loss is calculated as per equation (9).
0 20 40 60 80 100 120
1.5
2
2.5
3
3.5
4
4.5
5
x 10
-4
Months
T
i
m
e
-
a
v
e
r
a
g
e
d

v
a
r
i
a
n
c
e
0 20 40 60 80 100 120
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
Months
P
r
o
b
a
b
i
l
i
t
y

o
f

L
o
s
s
Time-averaged variance
Probability of loss
Fig 6. Probability of a loss and time-averaged variance for the value strategy CP/5 which represents a CM statistical arbitrage
opportunity. Sample period is from January 1965 to December 2000. The value strategy CP/5 longs the bottom past year cash flow to
price decile of stocks and shorts the top past year cash flow to price decile of stocks in the sample and holds the spread for 5 years. $1
is invested at all times. Any profits (losses) are re-invested in (financed with) the risk free asset. The probability of loss is calculated
as per equation (9).
0 20 40 60 80 100 120
0.002
0.004
0.006
0.008
0.01
0.012
0.014
0.016
0.018
Months
T
i
m
e
-
a
v
e
r
a
g
e
d

v
a
r
i
a
n
c
e
0 20 40 60 80 100 120
0.2
0.25
0.3
0.35
0.4
0.45
0.5
Months
P
r
o
b
a
b
i
l
i
t
y

o
f

L
o
s
s
Probability of loss
Time-averaged variance
Fig 7. Probability of a loss and time-averaged variance for the value strategy Sales/5 which represents a CM statistical arbitrage
opportunity. Sample period is from January 1965 to December 2000. The value strategy Sales/5 longs the bottom past 3-year sales
growth decile of stocks and shorts the top past 3-year sales growth decile of stocks in the sample and holds the spread for 5 years. $1
is invested at all times. Any profits (losses) are re-invested in (financed with) the risk free asset. The probability of loss is calculated
as per equation (9).
0 20 40 60 80 100 120
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
x 10
-3
Months
T
i
m
e
-
a
v
e
r
a
g
e
d

v
a
r
i
a
n
c
e
0 20 40 60 80 100 120
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
Months
P
r
o
b
a
b
i
l
i
t
y

o
f

L
o
s
s
Time-averaged variance
Probability of loss
Fig 8. Cumulative Trading Profits from $1 invested in the Fama and French (1992) SMB portfolio. Sample period is from January 1965 to
December 2000. The SMB portfolio longs stocks with market equity (ME) below the 30th NYSE ME percentile and shorts stocks above the
70th NYSE ME percentile. Any profits (losses) are re-invested in (financed with) the risk free asset.
Jan 1970 Jan 1980 Jan 1990 Jan 2000
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Time (months)
C
u
m
u
l
a
t
i
v
e

p
r
o
f
i
t

i
n

d
o
l
l
a
r
s
Min Max Median Mean Std. Dev.
MOM 3/3 -0.1931 0.1176 0.0005 -0.0026 0.0301
MOM 3/6 -0.0939 0.095 0.003 0.0016 0.0192
MOM 3/9 -0.0676 0.0544 0.0037 0.0022 0.015
MOM 3/12 -0.054 0.0398 0.0033 0.0034 0.0121
MOM 6/3 -0.2042 0.1623 0.0049 0.0021 0.0339
MOM 6/6 -0.0994 0.0524 0.0055 0.0043 0.02
MOM 6/9 -0.067 0.0513 0.007 0.0055 0.0166
MOM 6/12 -0.0576 0.0475 0.0043 0.0044 0.0129
MOM 9/3 -0.2104 0.1095 0.0066 0.0036 0.0327
MOM 9/6 -0.1063 0.0621 0.0074 0.0065 0.0203
MOM 9/9 -0.0611 0.054 0.0064 0.0053 0.0157
MOM 9/12 -0.0446 0.0378 0.0041 0.0033 0.0117
MOM 12/3 -0.1596 0.1165 0.008 0.0061 0.0315
MOM 12/6 -0.0803 0.0718 0.0063 0.0057 0.0189
MOM 12/9 -0.0635 0.0539 0.0048 0.0037 0.0145
MOM 12/12 -0.0331 0.0315 0.0019 0.0015 0.0104
BM1 -0.1469 0.2719 0.0126 0.0139 0.0445
BM3 -0.0997 0.1768 0.0126 0.0115 0.0337
BM5 -0.0843 0.1426 0.0102 0.0104 0.0291
CP1 -0.3524 0.5473 0.0092 0.0045 0.0722
CP3 -0.239 0.2169 0.0067 0.0042 0.0462
CP5 -0.2139 0.1229 0.0065 0.0032 0.0409
EP1 -0.3683 0.1578 0.0059 0.0006 0.0515
EP3 -0.2785 0.1169 0.0037 0.0011 0.0379
EP5 -0.188 0.0715 0.0032 0.0011 0.031
SALES1 -0.065 0.1607 0.0076 0.0085 0.0285
SALES3 -0.0564 0.1022 0.0062 0.0058 0.0229
SALES5 -0.0777 0.0789 0.0047 0.0048 0.0212
Table 1
Summary Statistics for the Incremental Profits from Momentum and Value
Strategies
Summary statistics for momentum and value portfolios. Sample period is from January 1965 to
December 2000. Portfolio MOM x/y denotes a momentum portfolio with a formation period of x
months and a holding period of y months. Every month, stocks are sorted based on past x months
of returns into decile portfolios. The portfolio MOM x/y longs the top decile and shorts the
bottom decile and holds that spread for y months as in Jegadeesh and Titman (1993). The other
portfolios are value portfolios. As in Lakonishok, Shleifer, and Vishny (1994), the sorting
variables are past one-year book-to-market (BM), cash flow-to-price (CP), and earnings-to-price
(EP) as well as past three-year sales growth (SALES). Portfolio xy denotes a value portfolio based
on stock characteristic x and a holding period of y years. Every July, stocks are sorted into decile
portfolios based on a particular stock characteristic. For BM, CP, and EP portfolios, the portfolio
xy longs the top decile and shorts the bottom decile and holds that spread for y years. For SALES
portfolios, the portfolio xy longs the bottom decile and shorts the top decile and holds that spread
for y years. The risk free asset is used to finance the portfolios.
Portfolio (mean) t-stat (std dev) (growth rate H1 (>0) H2 (<0) Sum (H1+H2) Statistical Arbitrage
of std dev)
MOM 3/3 -0.0027 -1.78 0.033 -0.0174 0.967 0.306 1.273 No
MOM 3/6^ 0.0015 1.68 0.023 -0.0350 0.056 0.156 0.212 No
MOM 3/9^ 0.0021 3.08 0.025 -0.1083 0.001 0.001 0.002 Yes**
MOM 3/12^ 0.0033 5.81 0.023 -0.1317 0.000 0.000 0.000 Yes**
MOM 6/3^ 0.002 1.27 0.037 -0.0168 0.11 0.306 0.416 No
MOM 6/6^ 0.0042 4.45 0.037 -0.1254 0.000 0.000 0.000 Yes**
MOM 6/9^ 0.0055 6.78 0.026 -0.0908 0.000 0.004 0.004 Yes**
MOM 6/12^ 0.0042 6.92 0.03 -0.1734 0.000 0.000 0.000 Yes**
MOM 9/3^ 0.0035 2.24 0.043 -0.0532 0.013 0.075 0.088 Yes*
MOM 9/6^ 0.0065 6.53 0.027 -0.0603 0.000 0.052 0.052 Yes*
MOM 9/9^ 0.0053 6.93 0.021 -0.0549 0.000 0.064 0.064 Yes*
MOM 9/12^ 0.0031 5.78 0.022 -0.1256 0.000 0.001 0.001 Yes**
MOM 12/3^ 0.006 3.93 0.035 -0.0206 0.000 0.266 0.266 No
MOM 12/6^ 0.0057 6.13 0.02 -0.0100 0.000 0.385 0.385 No
MOM 12/9 0.0038 5.19 0.012 0.0432 0.000 0.878 0.878 No
MOM 12/12^ 0.0015 3.00 0.01 -0.0004 0.002 0.496 0.498 No
* Significant at the 10% level
** Significant at the 5% level
^ Point estimates of and consistent with statistical arbitrage
Parameter estimates and corresponding p values for constrained mean test of statistical arbitrage. Sample period is from January 1965 to
December 2000. Portfolio x/y denotes a momentum portfolio with a formation period of x months and a holding period of y months. Every
month, stocks are sorted based on past x months of returns into decile portfolios. The portfolio x/y longs the top decile and shorts the bottom
decile and holds that spread for y months as in Jegadeesh and Titman (1993). The risk free asset is used to finance the portfolio. H1 and H2
denote the p values from statistical arbitrage tests which test whether the portfolio's mean monthly incremental profit is positive and whether
its time-averaged variance is declining over time. The sum of the H1 and H2 columns is the p value for the statistical arbitrage test. Note that
the sum may exceed one because of the Bonferonni inequality. The t-statistic on the mean monthly trading profit is provided for comparison.
Table 2
Constrained Mean (CM) Statistical Arbitrage Tests: Jegadeesh and Titman (1993) Momentum Strategies
Portfolio (mean) t-stat (std dev) (growth rate H1 (>0) H2 (<0) Sum (H1+H2) Statistical Arbitrage
of std dev)
BM1^ 0.0141 6.34 0.054 -0.0373 0.000 0.174 0.174 No
BM3^ 0.0115 6.54 0.042 -0.0427 0.000 0.166 0.166 No
BM5^ 0.0106 6.46 0.041 -0.0720 0.000 0.063 0.063 Yes*
CP1^ 0.0021 1.28 0.196 -0.2179 0.249 0.000 0.249 No
CP3^ 0.0036 1.74 0.132 -0.2272 0.046 0.000 0.046 Yes**
CP5^ 0.0037 1.39 0.131 -0.2685 0.027 0.000 0.027 Yes**
EP1 0.0021 0.25 0.015 0.2320 0.177 1.000 1.177 No
EP3 0.0013 0.57 0.026 0.0735 0.246 0.980 1.226 No
EP5^ 0.0011 0.63 0.039 -0.0494 0.268 0.079 0.347 No
SALES1^ 0.0085 5.79 0.043 -0.0856 0.000 0.019 0.019 Yes**
SALES3^ 0.0056 4.65 0.036 -0.0936 0.000 0.028 0.028 Yes**
SALES5^ 0.0046 3.83 0.042 -0.1517 0.000 0.001 0.001 Yes**
* Significant at the 10% level
** Significant at the 5% level
^ Point estimates of and consistent with statistical arbitrage
Table 3
Constrained Mean (CM) Statistical Arbitrage Tests: Lakonishok, Shleifer, and Vishny (1994) Value Strategies
Parameter estimates and corresponding p values for constrained mean test of statistical arbitrage. Sample period is from January 1965 to
December 2000. As in Lakonishok, Shleifer, and Vishny (1994), the sorting variables are past one-year book-to-market (BM), cash flow-to-
price (CP), and earnings-to-price (EP) as well as past three-year sales growth (SALES). Portfolio xy denotes a value portfolio based on stock
characteristic x and a holding period of y years. Every July, stocks are sorted into decile portfolios based on a particular stock characteristic.
For BM, CP, and EP portfolios, the portfolio xy longs the top decile and shorts the bottom decile and holds that spread for y years. For
SALES portfolios, the portfolio xy longs the bottom decile and shorts the top decile and holds that spread for y years. The risk free asset is
used to finance the portfolio. H1 and H2 denote the p values from statistical arbitrage tests which test whether the portfolio's mean monthly
incremental profit is positive and whether its time-averaged variance is declining over time. The sum of the H1 and H2 columns is the p value
for the statistical arbitrage test. Note that the sum may exceed one because of the Bonferonni inequality. The t-statistic on the mean monthly
trading profit is provided for comparison.
Portfolio RMSE RMSE SSR SSR p-value, (growth Likelihood Ratio
(CM) (UM) (CM) (UM) rate of mean) <0 Test
MOM 3/3 0.879 0.877 1.000 1.000 0.83 0.037
MOM 3/6 0.558 0.556 1.001 1.000 0.00 0.066
MOM 3/9** 0.433 0.432 1.003 1.003 0.14 0.017
MOM 3/12** 0.35 0.349 1.004 1.004 0.20 0.014
MOM 6/3 0.986 0.98 1.000 1.000 0.00 0.150
MOM 6/6** 0.581 0.58 1.003 1.003 0.38 0.007
MOM 6/9** 0.479 0.479 1.002 1.002 0.41 0.002
MOM 6/12** 0.371 0.37 1.005 1.004 0.25 0.012
MOM 9/3* 0.948 0.947 1.001 1.001 0.17 0.020
MOM 9/6* 0.587 0.587 1.002 1.002 0.53 0.000
MOM 9/9* 0.45 0.45 1.001 1.001 0.32 0.004
MOM 9/12** 0.335 0.333 1.003 1.003 0.01 0.049
MOM 12/3 0.909 0.907 1.001 1.001 0.07 0.040
MOM 12/6 0.542 0.542 1.000 1.000 0.14 0.017
MOM 12/9 0.417 0.415 0.999 0.999 0.01 0.047
MOM 12/12 0.297 0.293 1.000 1.000 0.00 0.124
BM1 1.279 1.277 1.001 1.001 0.79 0.044
BM3 0.919 0.919 1.001 1.001 0.69 0.010
BM5* 0.74 0.74 1.002 1.002 0.69 0.014
CP1 2.04 2.03 1.009 1.008 0.00 0.206
CP3** 1.255 1.255 1.007 1.007 0.20 0.013
CP5** 1.024 1.024 1.012 1.012 0.55 0.003
EP1 1.468 1.468 0.991 0.997 0.05 0.078
EP3 1.032 1.03 0.998 0.998 0.24 0.012
EP5 0.788 0.788 1.002 1.002 0.43 0.000
SALES1** 0.782 0.782 1.002 1.002 0.48 0.000
SALES3** 0.593 0.592 1.002 1.002 0.19 0.011
SALES5** 0.509 0.509 1.005 1.005 0.28 0.006
* Constrained Mean Statistical Arbitrages at the 10% level
** Constrained Mean Statistical Arbitrages at the 5% level
Table 4
Comparison Between Constrained Mean and Unconstrained Mean Models
A comparison of the Root Mean Squared Errors (RMSE) and Sum of Squared Residuals (SSR)
between constrained mean and unconstrained mean models of statistical arbitrage for the
momentum portfolios and value portfolios in Tables 2 and 3. The RMSE is based on a monte carlo
experiment with 10,000 simulated incremental trading profit time series. The p-value for the t-ratio
test that the incremental trading profit of the portfolios is declining over time is also presented. The
likelihood ratio test evaluates the null that the rate of change of the mean is zero. The test statistics
associated with the likelihood ratio tests are displayed and would have to exceed the critical value of
2.71 in order to reject the null at the 10% level (see footnote 16).
Type of deviation
at the 5% level at the 10% level
No Deviation 99% 100%
Autocorrelation 95% 98%
Jumps 60% 70%
Non Stationary , 71% 81%
Type of deviation
at the 5% level at the 10% level
No Deviation 100% 95%
Autocorrelation 100% 97%
Jumps 99% 95%
Non Stationary , 100% 95%
Panel B: True State = No Statistical Arbitrage
Percentage of simulations which correctly accept the null of no statistical arbitrage
Percentage of simulations which correctly reject the null of no statistical arbitrage
Table 5
Simulation Accuracy Tests With Deviations From Assumed Process
This table reports simulation results on the accuracy of the statistical arbitrage test in the face of deviations from the assumed
incremental profit process. The deviations include autocorrelation, jumps, and parameter non stationarity. Please refer to Section 6.2
for the exact specifications of the various deviations. Panel A details the situations when the original parameters are consistent with
statistical arbitrage ( is set to -0.15). Panel B details the situations when the original parameters are consistent with no statistical
arbitrage ( is set to 0). The first column details the departure from the assumed dynamics in equation (1). Columns two and three
record the number of instances in the 10,000 simulated time series that the null hypothesis of no statistical arbitrage is correctly
rejected or accepted at the 5% and the 10% levels.
Panel A : True State = Statistical Arbitrage
Portfolio Turnover (%) Max loss Max (growth H1 (>0) H2 (<0) Sum (H1+H2) Statistical
of $1 invested drawdown (mean) (std dev) rate of std dev) Arbitrage
MOM 3/9^ 9.26 0.0949 0.0676 0.0006 0.0232 -0.108 0.193 0.001 0.195 No
MOM 3/12^ 7.22 0.0425 0.0540 0.0019 0.0210 -0.131 0.000 0.000 0.000 Yes**
MOM 6/6^ 14.20 0.1131 0.0994 0.0019 0.0337 -0.125 0.016 0.000 0.016 Yes**
MOM 6/9^ 9.49 0.1091 0.0670 0.0036 0.0236 -0.091 0.000 0.004 0.004 Yes**
MOM 6/12^ 7.23 0.0418 0.0576 0.0027 0.0271 -0.173 0.000 0.000 0.000 Yes**
MOM 9/12^ 7.24 0.0052 0.0446 0.0017 0.0196 -0.125 0.000 0.001 0.001 Yes**
CP3^ 2.17 0.2250 0.2390 0.0028 0.1198 -0.227 0.075 0.000 0.075 Yes*
CP5^ 1.46 0.3957 0.2139 0.0029 0.1189 -0.268 0.045 0.000 0.045 Yes**
SALES1^ 4.07 0.0817 0.0650 0.0070 0.0392 -0.086 0.000 0.019 0.019 Yes**
SALES3^ 2.41 0.0246 0.0564 0.0046 0.0324 -0.094 0.000 0.028 0.028 Yes**
SALES5^ 1.52 0.0487 0.0777 0.0038 0.0383 -0.152 0.000 0.001 0.001 Yes**
^ Point estimates of and consistent with statistical arbitrage
* Significant at the 10% level, ** Significant at the 5% level
Table 6
Transactions Costs and Other Market Frictions Adjusted, Constrained Mean (CM) Tests of Statistical Arbitrage
Constrained test of statistical arbitrage after adjusting for market frictions like transactions cost, margins, liquidity buffers, haircuts, and a higher
borrowing rate. The sample period and column variables are as per defined in Tables 2 and 3. All parameter estimates are calculated from portfolio
incremental profits after adjusting for five types of market frictions: (I) margin (m) required on long and short positions, (II) liquidity buffer (c) for
short positions, (III) the haircut (h) implicity levied on the margin interest, (IV) a higher borrowing rate than lending rate, and (V) transactions costs.
The values of h and c are set at 27.5 basis points and 10% respectively, following Jacobs and Levy (1995, 1997). The value of m is set at a
conservative 50%. The borrowing rate is set 2% higher than the lending rate (which equals the risk free rate). This spread is higher than the average
spread between the AA 30 day commercial paper rate and the risk free rate from 1997 - 2002. The incremental profits are adjusted for (I), (II), (III),
and (IV) using the methodology of Alexander (2000) and the adjustment process is detailed in equation (10) of the text. Estimated transactions costs p
cost per month is estimated round-trip transactions cost x turnover where round-trip transactions cost = 1.34%, the transactions costs for the average
stock traded in the NASDAQ (Chan and Lakonishok, 1997). Turnover is the number of round trip transactions per month over the number of stocks
held in the portfolio. Max loss is the minimum cumulative profit while max drawdown is the minimum monthly incremental profit.